{ "cells": [ { "cell_type": "code", "execution_count": 1, "id": "fecc3e39", "metadata": {}, "outputs": [], "source": [ "from dotenv import load_dotenv\n", "load_dotenv(\".env\", override=True)\n", "%load_ext autoreload\n", "%autoreload 2" ] }, { "cell_type": "markdown", "id": "4c897bc9", "metadata": {}, "source": [ "# Deepagent for Research\n", "\n", "We will use the `deepagents` package to create a research agent. When using the `deepagents` package, it's important to: \n", "\n", "1. Understand the native tools available\n", "2. Supply task-specific tools\n", "3. Supply task-specific instructions\n", "4. Supply task-specific sub-agents" ] }, { "cell_type": "markdown", "id": "3a14c09e", "metadata": {}, "source": [ "## Task-Specific Tools \n", "\n", "You can see an overview of the native tools in the [deepagents package README](https://github.com/langchain-ai/deepagents?tab=readme-ov-file#model) as well as the [quickstarts README](https://github.com/langchain-ai/deepagents-quickstarts). We'll extend this with two task-specific tools. \n", "\n", "### Search Tool \n", "\n", "There are different search tools that we can use. For example, we can use [Tavily](https://www.tavily.com/) to search for relevant URLs, then fetches the full webpage content.\n", "\n", "### Think Tool \n", "\n", "We'll supply a [think tool](https://www.anthropic.com/engineering/claude-think-tool), which is a useful way to help audit agent decision making. " ] }, { "cell_type": "code", "execution_count": 2, "id": "9163556f", "metadata": {}, "outputs": [], "source": [ "from research_agent.tools import tavily_search, think_tool\n", "tools = [tavily_search, think_tool]" ] }, { "cell_type": "markdown", "id": "7ba3ee5b", "metadata": {}, "source": [ "## Task-Specific Instructions\n", " \n", "Next, let's define task specific instructions using [a few prompting techniques for agents](https://youtu.be/XSZP9GhhuAc?si=zowpViL-2j-vI9hA):\n", "\n", "### 1. Think Like The Agent\n", "What instructions would you give a new work colleague?\n", "- **Read the question carefully** - What specific information does the user need?\n", "- **Start with broader searches** - Use broad, comprehensive queries first\n", "- **After each search, pause and assess** - Do I have enough to answer? What's still missing?\n", "- **Execute narrower searches as you gather information** - Fill in the gaps.\n", "\n", "### 2. Concrete Heuristics (Prevent \"Spin-Out\" on excessive tool calls)\n", "Use **Hard Limits** to prevent the research agent from calling tools excessively:\n", "- **Stop when you can answer confidently** - Don't keep searching for perfection.\n", "- **Give it budgets** - Use 2-3 search tool calls for simple queries. Use up to 5 for complex queries.\n", "- **Limit** - Always stop after 5 search tool calls if you cannot find the right source(s).\n", "\n", "### 3. Show your thinking\n", "After each search tool calling, use [`think_tool` to analyze the results](https://www.anthropic.com/engineering/claude-think-tool):\n", "- What key information did I find? \n", "- What's missing?\n", "- Do I have enough to answer the question comprehensively?\n", "- Should I search more or provide my answer?" ] }, { "cell_type": "code", "execution_count": 3, "id": "4487f04d", "metadata": {}, "outputs": [], "source": [ "from datetime import datetime\n", "from utils import show_prompt, format_messages\n", "from research_agent.prompts import (\n", " RESEARCHER_INSTRUCTIONS,\n", " RESEARCH_WORKFLOW_INSTRUCTIONS,\n", " SUBAGENT_DELEGATION_INSTRUCTIONS,\n", ")" ] }, { "cell_type": "code", "execution_count": 4, "id": "a5eb7a89-8a26-4fb4-ba77-b05180f2c67e", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
╭──────────────────────────────────────────────────── Prompt ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│  You are a research assistant conducting research on the user's input topic. For context, today's date is       │\n",
       "│  {date}.                                                                                                        │\n",
       "│                                                                                                                 │\n",
       "│  <Task>                                                                                                         │\n",
       "│  Your job is to use tools to gather information about the user's input topic.                                   │\n",
       "│  You can use any of the research tools provided to you to find resources that can help answer the research      │\n",
       "│  question.                                                                                                      │\n",
       "│  You can call these tools in series or in parallel, your research is conducted in a tool-calling loop.          │\n",
       "│  </Task>                                                                                                        │\n",
       "│                                                                                                                 │\n",
       "│  <Available Research Tools>                                                                                     │\n",
       "│  You have access to two specific research tools:                                                                │\n",
       "│  1. **tavily_search**: For conducting web searches to gather information                                        │\n",
       "│  2. **think_tool**: For reflection and strategic planning during research                                       │\n",
       "│  **CRITICAL: Use think_tool after each search to reflect on results and plan next steps**                       │\n",
       "│  </Available Research Tools>                                                                                    │\n",
       "│                                                                                                                 │\n",
       "│  <Instructions>                                                                                                 │\n",
       "│  Think like a human researcher with limited time. Follow these steps:                                           │\n",
       "│                                                                                                                 │\n",
       "│  1. **Read the question carefully** - What specific information does the user need?                             │\n",
       "│  2. **Start with broader searches** - Use broad, comprehensive queries first                                    │\n",
       "│  3. **After each search, pause and assess** - Do I have enough to answer? What's still missing?                 │\n",
       "│  4. **Execute narrower searches as you gather information** - Fill in the gaps                                  │\n",
       "│  5. **Stop when you can answer confidently** - Don't keep searching for perfection                              │\n",
       "│  </Instructions>                                                                                                │\n",
       "│                                                                                                                 │\n",
       "│  <Hard Limits>                                                                                                  │\n",
       "│  **Tool Call Budgets** (Prevent excessive searching):                                                           │\n",
       "│  - **Simple queries**: Use 2-3 search tool calls maximum                                                        │\n",
       "│  - **Complex queries**: Use up to 5 search tool calls maximum                                                   │\n",
       "│  - **Always stop**: After 5 search tool calls if you cannot find the right sources                              │\n",
       "│                                                                                                                 │\n",
       "│  **Stop Immediately When**:                                                                                     │\n",
       "│  - You can answer the user's question comprehensively                                                           │\n",
       "│  - You have 3+ relevant examples/sources for the question                                                       │\n",
       "│  - Your last 2 searches returned similar information                                                            │\n",
       "│  </Hard Limits>                                                                                                 │\n",
       "│                                                                                                                 │\n",
       "│  <Show Your Thinking>                                                                                           │\n",
       "│  After each search tool call, use think_tool to analyze the results:                                            │\n",
       "│  - What key information did I find?                                                                             │\n",
       "│  - What's missing?                                                                                              │\n",
       "│  - Do I have enough to answer the question comprehensively?                                                     │\n",
       "│  - Should I search more or provide my answer?                                                                   │\n",
       "│  </Show Your Thinking>                                                                                          │\n",
       "│                                                                                                                 │\n",
       "│  <Final Response Format>                                                                                        │\n",
       "│  When providing your findings back to the orchestrator:                                                         │\n",
       "│                                                                                                                 │\n",
       "│  1. **Structure your response**: Organize findings with clear headings and detailed explanations                │\n",
       "│  2. **Cite sources inline**: Use [1], [2], [3] format when referencing information from your searches           │\n",
       "│  3. **Include Sources section**: End with ### Sources listing each numbered source with title and URL           │\n",
       "│                                                                                                                 │\n",
       "│  Example:                                                                                                       │\n",
       "│  ```                                                                                                            │\n",
       "│  ## Key Findings                                                                                                │\n",
       "│                                                                                                                 │\n",
       "│  Context engineering is a critical technique for AI agents [1]. Studies show that proper context management     │\n",
       "│  can improve performance by 40% [2].                                                                            │\n",
       "│                                                                                                                 │\n",
       "│  ### Sources                                                                                                    │\n",
       "│  [1] Context Engineering Guide: https://example.com/context-guide                                               │\n",
       "│  [2] AI Performance Study: https://example.com/study                                                            │\n",
       "│  ```                                                                                                            │\n",
       "│                                                                                                                 │\n",
       "│  The orchestrator will consolidate citations from all sub-agents into the final report.                         │\n",
       "│  </Final Response Format>                                                                                       │\n",
       "│                                                                                                                 │\n",
       "│                                                                                                                 │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[34m╭─\u001b[0m\u001b[34m───────────────────────────────────────────────────\u001b[0m\u001b[34m \u001b[0m\u001b[1;32mPrompt\u001b[0m\u001b[34m \u001b[0m\u001b[34m────────────────────────────────────────────────────\u001b[0m\u001b[34m─╮\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m You are a research assistant conducting research on the user's input topic. For context, today's date is \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m {date}. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Your job is to use tools to gather information about the user's input topic. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m You can use any of the research tools provided to you to find resources that can help answer the research \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m question. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m You can call these tools in series or in parallel, your research is conducted in a tool-calling loop. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m You have access to two specific research tools: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. **tavily_search**: For conducting web searches to gather information \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. **think_tool**: For reflection and strategic planning during research \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **CRITICAL: Use think_tool after each search to reflect on results and plan next steps** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Think like a human researcher with limited time. Follow these steps: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. **Read the question carefully** - What specific information does the user need? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. **Start with broader searches** - Use broad, comprehensive queries first \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. **After each search, pause and assess** - Do I have enough to answer? What's still missing? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 4. **Execute narrower searches as you gather information** - Fill in the gaps \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 5. **Stop when you can answer confidently** - Don't keep searching for perfection \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **Tool Call Budgets** (Prevent excessive searching): \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Simple queries**: Use 2-3 search tool calls maximum \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Complex queries**: Use up to 5 search tool calls maximum \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Always stop**: After 5 search tool calls if you cannot find the right sources \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **Stop Immediately When**: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - You can answer the user's question comprehensively \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - You have 3+ relevant examples/sources for the question \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Your last 2 searches returned similar information \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m After each search tool call, use think_tool to analyze the results: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - What key information did I find? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - What's missing? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Do I have enough to answer the question comprehensively? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Should I search more or provide my answer? \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m When providing your findings back to the orchestrator: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. **Structure your response**: Organize findings with clear headings and detailed explanations \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. **Cite sources inline**: Use [1], [2], [3] format when referencing information from your searches \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. **Include Sources section**: End with \u001b[1;36m#\u001b[0m\u001b[1;36m## Sources listing each numbered source with title and URL\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Example: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m ``` \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Key Findings\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Context engineering is a critical technique for AI agents [1]. Studies show that proper context management \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m can improve performance by 40% [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Sources\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [1] Context Engineering Guide: https://example.com/context-guide \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [2] AI Performance Study: https://example.com/study \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m ``` \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m The orchestrator will consolidate citations from all sub-agents into the final report. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;34m\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "show_prompt(RESEARCHER_INSTRUCTIONS)" ] }, { "cell_type": "markdown", "id": "48ab6e7e", "metadata": {}, "source": [ "## Task-Specific Sub-Agents\n", "\n", "You can specify [custom subagents](https://github.com/langchain-ai/deepagents?tab=readme-ov-file#subagents) as a means of context isolation. \n", "\n", "Here's well define a sub-agent that can search the web for information. " ] }, { "cell_type": "code", "execution_count": 5, "id": "a6570183", "metadata": {}, "outputs": [], "source": [ "# Get current date\n", "current_date = datetime.now().strftime(\"%Y-%m-%d\")\n", "\n", "# Create research sub-agent\n", "research_sub_agent = {\n", " \"name\": \"research-agent\",\n", " \"description\": \"Delegate research to the sub-agent researcher. Only give this researcher one topic at a time.\",\n", " \"system_prompt\": RESEARCHER_INSTRUCTIONS.format(date=current_date),\n", " \"tools\": [tavily_search, think_tool],\n", "}" ] }, { "cell_type": "markdown", "id": "8ef99f1d", "metadata": {}, "source": [ "## Putting it all together\n", "\n", "### Instructions\n", "\n", "Now, we can look at all of our instructions together. " ] }, { "cell_type": "code", "execution_count": 6, "id": "4e55b2c0", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
╭──────────────────────────────────────────────────── Prompt ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│  # Research Workflow                                                                                            │\n",
       "│                                                                                                                 │\n",
       "│  Follow this workflow for all research requests:                                                                │\n",
       "│                                                                                                                 │\n",
       "│  1. **Plan**: Create a todo list with write_todos to break down the research into focused tasks                 │\n",
       "│  2. **Save the request**: Use write_file() to save the user's research question to `/research_request.md`       │\n",
       "│  3. **Research**: Delegate research tasks to sub-agents using the task() tool - ALWAYS use sub-agents for       │\n",
       "│  research, never conduct research yourself                                                                      │\n",
       "│  4. **Synthesize**: Review all sub-agent findings and consolidate citations (each unique URL gets one number    │\n",
       "│  across all findings)                                                                                           │\n",
       "│  5. **Write Report**: Write a comprehensive final report to `/final_report.md` (see Report Writing Guidelines   │\n",
       "│  below)                                                                                                         │\n",
       "│  6. **Verify**: Read `/research_request.md` and confirm you've addressed all aspects with proper citations and  │\n",
       "│  structure                                                                                                      │\n",
       "│                                                                                                                 │\n",
       "│  ## Research Planning Guidelines                                                                                │\n",
       "│  - Batch similar research tasks into a single TODO to minimize overhead                                         │\n",
       "│  - For simple fact-finding questions, use 1 sub-agent                                                           │\n",
       "│  - For comparisons or multi-faceted topics, delegate to multiple parallel sub-agents                            │\n",
       "│  - Each sub-agent should research one specific aspect and return findings                                       │\n",
       "│                                                                                                                 │\n",
       "│  ## Report Writing Guidelines                                                                                   │\n",
       "│                                                                                                                 │\n",
       "│  When writing the final report to `/final_report.md`, follow these structure patterns:                          │\n",
       "│                                                                                                                 │\n",
       "│  **For comparisons:**                                                                                           │\n",
       "│  1. Introduction                                                                                                │\n",
       "│  2. Overview of topic A                                                                                         │\n",
       "│  3. Overview of topic B                                                                                         │\n",
       "│  4. Detailed comparison                                                                                         │\n",
       "│  5. Conclusion                                                                                                  │\n",
       "│                                                                                                                 │\n",
       "│  **For lists/rankings:**                                                                                        │\n",
       "│  Simply list items with details - no introduction needed:                                                       │\n",
       "│  1. Item 1 with explanation                                                                                     │\n",
       "│  2. Item 2 with explanation                                                                                     │\n",
       "│  3. Item 3 with explanation                                                                                     │\n",
       "│                                                                                                                 │\n",
       "│  **For summaries/overviews:**                                                                                   │\n",
       "│  1. Overview of topic                                                                                           │\n",
       "│  2. Key concept 1                                                                                               │\n",
       "│  3. Key concept 2                                                                                               │\n",
       "│  4. Key concept 3                                                                                               │\n",
       "│  5. Conclusion                                                                                                  │\n",
       "│                                                                                                                 │\n",
       "│  **General guidelines:**                                                                                        │\n",
       "│  - Use clear section headings (## for sections, ### for subsections)                                            │\n",
       "│  - Write in paragraph form by default - be text-heavy, not just bullet points                                   │\n",
       "│  - Do NOT use self-referential language (\"I found...\", \"I researched...\")                                       │\n",
       "│  - Write as a professional report without meta-commentary                                                       │\n",
       "│  - Each section should be comprehensive and detailed                                                            │\n",
       "│  - Use bullet points only when listing is more appropriate than prose                                           │\n",
       "│                                                                                                                 │\n",
       "│  **Citation format:**                                                                                           │\n",
       "│  - Cite sources inline using [1], [2], [3] format                                                               │\n",
       "│  - Assign each unique URL a single citation number across ALL sub-agent findings                                │\n",
       "│  - End report with ### Sources section listing each numbered source                                             │\n",
       "│  - Number sources sequentially without gaps (1,2,3,4...)                                                        │\n",
       "│  - Format: [1] Source Title: URL (each on separate line for proper list rendering)                              │\n",
       "│  - Example:                                                                                                     │\n",
       "│                                                                                                                 │\n",
       "│    Some important finding [1]. Another key insight [2].                                                         │\n",
       "│                                                                                                                 │\n",
       "│    ### Sources                                                                                                  │\n",
       "│    [1] AI Research Paper: https://example.com/paper                                                             │\n",
       "│    [2] Industry Analysis: https://example.com/analysis                                                          │\n",
       "│                                                                                                                 │\n",
       "│                                                                                                                 │\n",
       "│  ================================================================================                               │\n",
       "│                                                                                                                 │\n",
       "│  # Sub-Agent Research Coordination                                                                              │\n",
       "│                                                                                                                 │\n",
       "│  Your role is to coordinate research by delegating tasks from your TODO list to specialized research            │\n",
       "│  sub-agents.                                                                                                    │\n",
       "│                                                                                                                 │\n",
       "│  ## Delegation Strategy                                                                                         │\n",
       "│                                                                                                                 │\n",
       "│  **DEFAULT: Start with 1 sub-agent** for most queries:                                                          │\n",
       "│  - \"What is quantum computing?\" → 1 sub-agent (general overview)                                                │\n",
       "│  - \"List the top 10 coffee shops in San Francisco\" → 1 sub-agent                                                │\n",
       "│  - \"Summarize the history of the internet\" → 1 sub-agent                                                        │\n",
       "│  - \"Research context engineering for AI agents\" → 1 sub-agent (covers all aspects)                              │\n",
       "│                                                                                                                 │\n",
       "│  **ONLY parallelize when the query EXPLICITLY requires comparison or has clearly independent aspects:**         │\n",
       "│                                                                                                                 │\n",
       "│  **Explicit comparisons** → 1 sub-agent per element:                                                            │\n",
       "│  - \"Compare OpenAI vs Anthropic vs DeepMind AI safety approaches\" → 3 parallel sub-agents                       │\n",
       "│  - \"Compare Python vs JavaScript for web development\" → 2 parallel sub-agents                                   │\n",
       "│                                                                                                                 │\n",
       "│  **Clearly separated aspects** → 1 sub-agent per aspect (use sparingly):                                        │\n",
       "│  - \"Research renewable energy adoption in Europe, Asia, and North America\" → 3 parallel sub-agents (geographic  │\n",
       "│  separation)                                                                                                    │\n",
       "│  - Only use this pattern when aspects cannot be covered efficiently by a single comprehensive search            │\n",
       "│                                                                                                                 │\n",
       "│  ## Key Principles                                                                                              │\n",
       "│  - **Bias towards single sub-agent**: One comprehensive research task is more token-efficient than multiple     │\n",
       "│  narrow ones                                                                                                    │\n",
       "│  - **Avoid premature decomposition**: Don't break \"research X\" into \"research X overview\", \"research X          │\n",
       "│  techniques\", \"research X applications\" - just use 1 sub-agent for all of X                                     │\n",
       "│  - **Parallelize only for clear comparisons**: Use multiple sub-agents when comparing distinct entities or      │\n",
       "│  geographically separated data                                                                                  │\n",
       "│                                                                                                                 │\n",
       "│  ## Parallel Execution Limits                                                                                   │\n",
       "│  - Use at most 3 parallel sub-agents per iteration                                                              │\n",
       "│  - Make multiple task() calls in a single response to enable parallel execution                                 │\n",
       "│  - Each sub-agent returns findings independently                                                                │\n",
       "│                                                                                                                 │\n",
       "│  ## Research Limits                                                                                             │\n",
       "│  - Stop after 3 delegation rounds if you haven't found adequate sources                                         │\n",
       "│  - Stop when you have sufficient information to answer comprehensively                                          │\n",
       "│  - Bias towards focused research over exhaustive exploration                                                    │\n",
       "│                                                                                                                 │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[34m╭─\u001b[0m\u001b[34m───────────────────────────────────────────────────\u001b[0m\u001b[34m \u001b[0m\u001b[1;32mPrompt\u001b[0m\u001b[34m \u001b[0m\u001b[34m────────────────────────────────────────────────────\u001b[0m\u001b[34m─╮\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m # Research Workflow \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Follow this workflow for all research requests: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. **Plan**: Create a todo list with write_todos to break down the research into focused tasks \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. **Save the request**: Use write_file() to save the user's research question to `/research_request.md` \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. **Research**: Delegate research tasks to sub-agents using the task() tool - ALWAYS use sub-agents for \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m research, never conduct research yourself \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 4. **Synthesize**: Review all sub-agent findings and consolidate citations (each unique URL gets one number \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m across all findings) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 5. **Write Report**: Write a comprehensive final report to `/final_report.md` (see Report Writing Guidelines \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m below) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 6. **Verify**: Read `/research_request.md` and confirm you've addressed all aspects with proper citations and \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m structure \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Research Planning Guidelines\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Batch similar research tasks into a single TODO to minimize overhead \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - For simple fact-finding questions, use 1 sub-agent \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - For comparisons or multi-faceted topics, delegate to multiple parallel sub-agents \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Each sub-agent should research one specific aspect and return findings \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Report Writing Guidelines\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m When writing the final report to `/final_report.md`, follow these structure patterns: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **For comparisons:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. Introduction \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. Overview of topic A \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. Overview of topic B \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 4. Detailed comparison \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 5. Conclusion \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **For lists/rankings:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Simply list items with details - no introduction needed: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. Item 1 with explanation \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. Item 2 with explanation \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. Item 3 with explanation \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **For summaries/overviews:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 1. Overview of topic \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 2. Key concept 1 \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 3. Key concept 2 \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 4. Key concept 3 \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 5. Conclusion \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **General guidelines:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Use clear section headings (\u001b[1;35m## for sections, \u001b[0m\u001b[1;36m#\u001b[0m\u001b[1;36m## for subsections)\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Write in paragraph form by default - be text-heavy, not just bullet points \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Do NOT use self-referential language (\"I found...\", \"I researched...\") \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Write as a professional report without meta-commentary \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Each section should be comprehensive and detailed \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Use bullet points only when listing is more appropriate than prose \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **Citation format:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Cite sources inline using [1], [2], [3] format \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Assign each unique URL a single citation number across ALL sub-agent findings \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - End report with \u001b[1;36m#\u001b[0m\u001b[1;36m## Sources section listing each numbered source\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Number sources sequentially without gaps (1,2,3,4...) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Format: [1] Source Title: URL (each on separate line for proper list rendering) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Example: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Some important finding [1]. Another key insight [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Sources\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [1] AI Research Paper: https://example.com/paper \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [2] Industry Analysis: https://example.com/analysis \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m ================================================================================ \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m # Sub-Agent Research Coordination \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Your role is to coordinate research by delegating tasks from your TODO list to specialized research \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m sub-agents. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Delegation Strategy\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **DEFAULT: Start with 1 sub-agent** for most queries: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"What is quantum computing?\" → 1 sub-agent (general overview) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"List the top 10 coffee shops in San Francisco\" → 1 sub-agent \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"Summarize the history of the internet\" → 1 sub-agent \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"Research context engineering for AI agents\" → 1 sub-agent (covers all aspects) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **ONLY parallelize when the query EXPLICITLY requires comparison or has clearly independent aspects:** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **Explicit comparisons** → 1 sub-agent per element: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"Compare OpenAI vs Anthropic vs DeepMind AI safety approaches\" → 3 parallel sub-agents \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"Compare Python vs JavaScript for web development\" → 2 parallel sub-agents \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **Clearly separated aspects** → 1 sub-agent per aspect (use sparingly): \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - \"Research renewable energy adoption in Europe, Asia, and North America\" → 3 parallel sub-agents (geographic \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m separation) \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Only use this pattern when aspects cannot be covered efficiently by a single comprehensive search \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Key Principles\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Bias towards single sub-agent**: One comprehensive research task is more token-efficient than multiple \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m narrow ones \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Avoid premature decomposition**: Don't break \"research X\" into \"research X overview\", \"research X \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m techniques\", \"research X applications\" - just use 1 sub-agent for all of X \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - **Parallelize only for clear comparisons**: Use multiple sub-agents when comparing distinct entities or \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m geographically separated data \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Parallel Execution Limits\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Use at most 3 parallel sub-agents per iteration \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Make multiple task() calls in a single response to enable parallel execution \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Each sub-agent returns findings independently \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## Research Limits\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Stop after 3 delegation rounds if you haven't found adequate sources \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Stop when you have sufficient information to answer comprehensively \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m - Bias towards focused research over exhaustive exploration \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Limits\n", "max_concurrent_research_units = 3\n", "max_researcher_iterations = 3\n", "\n", "# Combine orchestrator instructions (RESEARCHER_INSTRUCTIONS only for sub-agents)\n", "INSTRUCTIONS = (\n", " RESEARCH_WORKFLOW_INSTRUCTIONS\n", " + \"\\n\\n\"\n", " + \"=\" * 80\n", " + \"\\n\\n\"\n", " + SUBAGENT_DELEGATION_INSTRUCTIONS.format(\n", " max_concurrent_research_units=max_concurrent_research_units,\n", " max_researcher_iterations=max_researcher_iterations,\n", " )\n", ")\n", "\n", "show_prompt(INSTRUCTIONS)" ] }, { "cell_type": "markdown", "id": "ab6e3639", "metadata": {}, "source": [ "### Create the agent\n", "\n", "Now, we create our deepagent with these components. " ] }, { "cell_type": "code", "execution_count": null, "id": "e979ff23-e36a-45b2-bd52-03cf4171f36c", "metadata": {}, "outputs": [], "source": [ "from IPython.display import Image, display\n", "from deepagents import create_deep_agent\n", "from langchain.chat_models import init_chat_model\n", "from langchain_google_genai import ChatGoogleGenerativeAI\n", "\n", "# Model Gemini 3 \n", "model = ChatGoogleGenerativeAI(model=\"gemini-3-pro-preview\")\n", "\n", "# Model Claude 4.5\n", "model = init_chat_model(model=\"anthropic:claude-sonnet-4-5-20250929\", temperature=0.0)" ] }, { "cell_type": "code", "execution_count": 10, "id": "62da8411", "metadata": {}, "outputs": [ { "data": { "image/png": "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", "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Create the agent\n", "agent = create_deep_agent(\n", " model=model,\n", " tools=tools, \n", " system_prompt=INSTRUCTIONS,\n", " subagents=[research_sub_agent],\n", " )\n", " \n", "# Show the agent\n", "display(Image(agent.get_graph().draw_mermaid_png()))" ] }, { "cell_type": "code", "execution_count": 12, "id": "613634c2", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
╭─────────────────────────────────────────────────── 🧑 Human ────────────────────────────────────────────────────╮\n",
       "│ research context engineering approaches used to build AI agents                                                 │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[34m╭─\u001b[0m\u001b[34m──────────────────────────────────────────────────\u001b[0m\u001b[34m 🧑 Human \u001b[0m\u001b[34m───────────────────────────────────────────────────\u001b[0m\u001b[34m─╮\u001b[0m\n", "\u001b[34m│\u001b[0m research context engineering approaches used to build AI agents \u001b[34m│\u001b[0m\n", "\u001b[34m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_todos                                                                                       │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"todos\": [                                                                                                    │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"in_progress\",                                                                                  │\n",
       "│       \"content\": \"Save research request to /research_request.md\"                                                │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Research context engineering approaches for AI agents using a sub-agent\"                      │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Synthesize findings and write final report to /final_report.md\"                               │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Verify report against original request\"                                                       │\n",
       "│     }                                                                                                           │\n",
       "│   ]                                                                                                             │\n",
       "│ }                                                                                                               │\n",
       "│    ID: 61820e98-a91b-4f74-ac2c-f443af11d389                                                                     │\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_file                                                                                        │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"file_path\": \"/research_request.md\",                                                                          │\n",
       "│   \"content\": \"research context engineering approaches used to build AI agents\"                                  │\n",
       "│ }                                                                                                               │\n",
       "│    ID: 3a5e0fbf-0181-4d89-9501-c5fac4b3cef2                                                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_todos \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"todos\": [ \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"in_progress\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Save research request to /research_request.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Research context engineering approaches for AI agents using a sub-agent\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Synthesize findings and write final report to /final_report.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Verify report against original request\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ] \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: 61820e98-a91b-4f74-ac2c-f443af11d389 \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_file \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"file_path\": \"/research_request.md\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"research context engineering approaches used to build AI agents\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: 3a5e0fbf-0181-4d89-9501-c5fac4b3cef2 \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'in_progress'},    │\n",
       "│ {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'pending'},    │\n",
       "│ {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'pending'}, {'content': │\n",
       "│ 'Verify report against original request', 'status': 'pending'}]                                                 │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'in_progress'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'pending'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'pending'}, {'content': \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 'Verify report against original request', 'status': 'pending'}] \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated file /research_request.md                                                                               │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated file /research_request.md \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_todos                                                                                       │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"todos\": [                                                                                                    │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Save research request to /research_request.md\"                                                │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"in_progress\",                                                                                  │\n",
       "│       \"content\": \"Research context engineering approaches for AI agents using a sub-agent\"                      │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Synthesize findings and write final report to /final_report.md\"                               │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Verify report against original request\"                                                       │\n",
       "│     }                                                                                                           │\n",
       "│   ]                                                                                                             │\n",
       "│ }                                                                                                               │\n",
       "│    ID: b317fa1b-6f81-47df-b0ca-aa5e50bd2b70                                                                     │\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: task                                                                                              │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"subagent_type\": \"research-agent\",                                                                            │\n",
       "│   \"description\": \"Research context engineering approaches used to build AI agents.\\nFocus on:\\n1. Definition    │\n",
       "│ and core concepts of context engineering.\\n2. Key techniques and strategies (e.g., dynamic context injection,   │\n",
       "│ RAG, memory management/history pruning, summarization, prompt structuring).\\n3. Architectures or frameworks     │\n",
       "│ often used (e.g., how LangChain or LlamaIndex handle context).\\n4. Approaches for handling context window       │\n",
       "│ limits vs. utilizing long-context models.\\n5. Specific patterns like \\\"Needle in a Haystack\\\" optimization or   │\n",
       "│ \\\"Context Stuffing\\\" vs \\\"Context Selection\\\".\\n\\nReturn a comprehensive summary of findings with citations.\"   │\n",
       "│ }                                                                                                               │\n",
       "│    ID: b682976a-b1fa-47a1-89be-153134ba28c7                                                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_todos \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"todos\": [ \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Save research request to /research_request.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"in_progress\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Research context engineering approaches for AI agents using a sub-agent\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Synthesize findings and write final report to /final_report.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Verify report against original request\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ] \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: b317fa1b-6f81-47df-b0ca-aa5e50bd2b70 \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: task \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"subagent_type\": \"research-agent\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"description\": \"Research context engineering approaches used to build AI agents.\\nFocus on:\\n1. Definition \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m and core concepts of context engineering.\\n2. Key techniques and strategies (e.g., dynamic context injection, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m RAG, memory management/history pruning, summarization, prompt structuring).\\n3. Architectures or frameworks \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m often used (e.g., how LangChain or LlamaIndex handle context).\\n4. Approaches for handling context window \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m limits vs. utilizing long-context models.\\n5. Specific patterns like \\\"Needle in a Haystack\\\" optimization or \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \\\"Context Stuffing\\\" vs \\\"Context Selection\\\".\\n\\nReturn a comprehensive summary of findings with citations.\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: b682976a-b1fa-47a1-89be-153134ba28c7 \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'},      │\n",
       "│ {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status':                │\n",
       "│ 'in_progress'}, {'content': 'Synthesize findings and write final report to /final_report.md', 'status':         │\n",
       "│ 'pending'}, {'content': 'Verify report against original request', 'status': 'pending'}]                         │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 'in_progress'}, {'content': 'Synthesize findings and write final report to /final_report.md', 'status': \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 'pending'}, {'content': 'Verify report against original request', 'status': 'pending'}] \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Context engineering is the systematic practice of designing, managing, and optimizing the information provided  │\n",
       "│ to AI models to ensure accurate, consistent, and context-aware behavior. It moves beyond simple prompt          │\n",
       "│ engineering by treating context as a dynamic resource that must be architected, managed, and validated.         │\n",
       "│                                                                                                                 │\n",
       "│ ### 1. Definition and Core Concepts                                                                             │\n",
       "│ **Context Engineering** is the \"process of designing, testing, and iterating on the contextual information      │\n",
       "│ provided to AI agents to shape their behavior and improve task performance\" [1]. Unlike a static prompt,        │\n",
       "│ context in an agentic system is a living state that changes with every user interaction, tool output, and       │\n",
       "│ internal reasoning step.                                                                                        │\n",
       "│                                                                                                                 │\n",
       "│ *   **Core Concepts**:                                                                                          │\n",
       "│     *   **Context Window**: The working memory limit of the LLM (e.g., 128k tokens). Managing this finite       │\n",
       "│ resource is the central challenge.                                                                              │\n",
       "│     *   **State Management**: Tracking the \"world state\" (user intent, task progress, tool results) across a    │\n",
       "│ multi-turn conversation.                                                                                        │\n",
       "│     *   **Context Pollution**: The degradation of model performance caused by irrelevant or contradictory       │\n",
       "│ information in the context [2].                                                                                 │\n",
       "│     *   **Context Continuity**: Ensuring the agent retains critical information (like user constraints)         │\n",
       "│ throughout a long interaction without \"forgetting\" them due to truncation.                                      │\n",
       "│                                                                                                                 │\n",
       "│ ### 2. Key Techniques and Strategies                                                                            │\n",
       "│ Effective context engineering employs a mix of retrieval, summarization, and structural strategies:             │\n",
       "│                                                                                                                 │\n",
       "│ *   **Dynamic Context Injection**: Instead of loading all data at once, agents \"pull\" information only when     │\n",
       "│ needed. For example, an agent might use a `search_tool` to inject real-time web results into the context only   │\n",
       "│ after identifying a knowledge gap [2].                                                                          │\n",
       "│ *   **Retrieval-Augmented Generation (RAG)**: A standard pattern for handling large knowledge bases. The agent  │\n",
       "│ retrieves strictly relevant chunks (e.g., top-5 matches) from a vector database to populate the context window, │\n",
       "│ rather than stuffing the entire document [2, 3].                                                                │\n",
       "│ *   **Memory Management**:                                                                                      │\n",
       "│     *   **Short-term Memory**: Stores immediate conversation history (e.g., last 10 turns).                     │\n",
       "│     *   **Long-term Memory**: Persists key facts (e.g., \"User is a vegetarian\") in an external database         │\n",
       "│ (Postgres, Vector Store) and retrieves them only when relevant [2].                                             │\n",
       "│     *   **Pruning/Trimming**: Algorithms that remove older, less relevant tokens to free up space while         │\n",
       "│ preserving recent interactions [2].                                                                             │\n",
       "│ *   **Summarization**: Compressing verbose tool outputs. Instead of feeding raw HTML from a scraped website, an │\n",
       "│ intermediate \"summarizer agent\" distills it into key points before passing it to the main agent [2].            │\n",
       "│ *   **Prompt Structuring**: Using structured formats (like JSON schemas or XML tags) within the context to      │\n",
       "│ clearly separate \"System Instructions,\" \"User Input,\" and \"Tool Outputs,\" preventing the model from confusing   │\n",
       "│ data with instructions.                                                                                         │\n",
       "│                                                                                                                 │\n",
       "│ ### 3. Architectures and Frameworks                                                                             │\n",
       "│ Two primary frameworks dominate the landscape, each with distinct approaches to context handling:               │\n",
       "│                                                                                                                 │\n",
       "│ | Feature | **LangChain** [3, 4] | **LlamaIndex** [3, 4] |                                                      │\n",
       "│ | :--- | :--- | :--- |                                                                                          │\n",
       "│ | **Primary Focus** | Flexible workflow & agent orchestration | Data indexing & advanced retrieval (RAG) |      │\n",
       "│ | **Memory Classes** | `ConversationBufferMemory` (raw history)<br>`ConversationSummaryMemory` (LLM-summarized  │\n",
       "│ history)<br>`ChatMessageHistory` (storage backend) | `ChatEngine` with modes:<br>`context` (standard            │\n",
       "│ RAG)<br>`condense_question` (rewrites queries based on history) |                                               │\n",
       "│ | **Context Strategy** | **Chains**: Sequential passing of context between steps.<br>**Agents**: Dynamic tool   │\n",
       "│ selection to populate context. | **Indices**: `VectorStoreIndex` (semantic search), `ListIndex` (sequential),   │\n",
       "│ `TreeIndex` (hierarchical summaries). |                                                                         │\n",
       "│ | **Optimization** | `Context Compression`: Removing irrelevant tokens from retrieved docs. | `Response         │\n",
       "│ Synthesizers`: <br>`compact` (stuffs max chunks into prompt)<br>`tree_summarize` (hierarchical reduction) |     │\n",
       "│                                                                                                                 │\n",
       "│ ### 4. Handling Context Limits vs. Long-Context Models                                                          │\n",
       "│ A critical architectural decision is whether to use RAG or rely on massive context windows (e.g., Gemini 1.5    │\n",
       "│ Pro's 2M tokens).                                                                                               │\n",
       "│                                                                                                                 │\n",
       "│ *   **RAG + Small Context**:                                                                                    │\n",
       "│     *   **Pros**: Significantly cheaper (fewer tokens processed), lower latency, and often higher accuracy for  │\n",
       "│ specific fact retrieval.                                                                                        │\n",
       "│     *   **Cons**: Can miss \"global\" insights that require reading the entire corpus at once.                    │\n",
       "│ *   **Long-Context Models (LCW)**:                                                                              │\n",
       "│     *   **Pros**: Can ingest entire books or codebases, enabling \"whole-document\" reasoning.                    │\n",
       "│     *   **Cons**: \"Lost in the Middle\" phenomenon where models forget information in the middle of a massive    │\n",
       "│ context.                                                                                                        │\n",
       "│     *   **Trade-off**: Benchmarks show that RAG systems often outperform LCW models in \"Needle in a Haystack\"   │\n",
       "│ tests for large corpora (2M+ tokens) because they filter noise before the model sees it [3].                    │\n",
       "│                                                                                                                 │\n",
       "│ ### 5. Specific Patterns                                                                                        │\n",
       "│ *   **Needle in a Haystack Optimization**:                                                                      │\n",
       "│     *   **The Problem**: LLMs struggle to find a specific fact (\"needle\") buried in a large amount of           │\n",
       "│ irrelevant text (\"haystack\").                                                                                   │\n",
       "│     *   **Optimization**: Place critical instructions or facts at the very **beginning** or very **end** of the │\n",
       "│ context window, as models pay most attention to these areas (Serial Position Effect).                           │\n",
       "│ *   **Context Stuffing vs. Context Selection**:                                                                 │\n",
       "│     *   **Context Stuffing**: Naively filling the window with as much data as possible. This often leads to     │\n",
       "│ *context distraction* and higher costs [2].                                                                     │\n",
       "│     *   **Context Selection**: Intentionally curating what enters the window.                                   │\n",
       "│         *   *Techniques*: Using **Node Postprocessors** (LlamaIndex) to filter retrieved nodes by keywords or   │\n",
       "│ similarity score before they reach the LLM [4].                                                                 │\n",
       "│         *   *Query Transformations*: Rewriting a user's vague query into a precise search term to select better │\n",
       "│ context (e.g., LangChain's `HyDE` or LlamaIndex's `Multi-step query decomposition`) [4].                        │\n",
       "│                                                                                                                 │\n",
       "│ ### Sources                                                                                                     │\n",
       "│ [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering                         │\n",
       "│ [2] Context Engineering Strategies (n8n):                                                                       │\n",
       "│ https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/                       │\n",
       "│ [3] RAG vs Long Context Window Study:                                                                           │\n",
       "│ https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs                             │\n",
       "│ [4] LangChain & LlamaIndex Comparison:                                                                          │\n",
       "│ https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- │\n",
       "│ in-large-2b5f3fabc360                                                                                           │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Context engineering is the systematic practice of designing, managing, and optimizing the information provided \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m to AI models to ensure accurate, consistent, and context-aware behavior. It moves beyond simple prompt \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m engineering by treating context as a dynamic resource that must be architected, managed, and validated. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### 1. Definition and Core Concepts \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m **Context Engineering** is the \"process of designing, testing, and iterating on the contextual information \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m provided to AI agents to shape their behavior and improve task performance\" [1]. Unlike a static prompt, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context in an agentic system is a living state that changes with every user interaction, tool output, and \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m internal reasoning step. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Core Concepts**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Window**: The working memory limit of the LLM (e.g., 128k tokens). Managing this finite \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m resource is the central challenge. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **State Management**: Tracking the \"world state\" (user intent, task progress, tool results) across a \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m multi-turn conversation. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Pollution**: The degradation of model performance caused by irrelevant or contradictory \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m information in the context [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Continuity**: Ensuring the agent retains critical information (like user constraints) \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m throughout a long interaction without \"forgetting\" them due to truncation. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### 2. Key Techniques and Strategies \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Effective context engineering employs a mix of retrieval, summarization, and structural strategies: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Dynamic Context Injection**: Instead of loading all data at once, agents \"pull\" information only when \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m needed. For example, an agent might use a `search_tool` to inject real-time web results into the context only \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m after identifying a knowledge gap [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Retrieval-Augmented Generation (RAG)**: A standard pattern for handling large knowledge bases. The agent \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m retrieves strictly relevant chunks (e.g., top-5 matches) from a vector database to populate the context window, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m rather than stuffing the entire document [2, 3]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Memory Management**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Short-term Memory**: Stores immediate conversation history (e.g., last 10 turns). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Long-term Memory**: Persists key facts (e.g., \"User is a vegetarian\") in an external database \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m (Postgres, Vector Store) and retrieves them only when relevant [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Pruning/Trimming**: Algorithms that remove older, less relevant tokens to free up space while \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m preserving recent interactions [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Summarization**: Compressing verbose tool outputs. Instead of feeding raw HTML from a scraped website, an \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m intermediate \"summarizer agent\" distills it into key points before passing it to the main agent [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Prompt Structuring**: Using structured formats (like JSON schemas or XML tags) within the context to \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m clearly separate \"System Instructions,\" \"User Input,\" and \"Tool Outputs,\" preventing the model from confusing \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m data with instructions. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### 3. Architectures and Frameworks \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Two primary frameworks dominate the landscape, each with distinct approaches to context handling: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | Feature | **LangChain** [3, 4] | **LlamaIndex** [3, 4] | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | :--- | :--- | :--- | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | **Primary Focus** | Flexible workflow & agent orchestration | Data indexing & advanced retrieval (RAG) | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | **Memory Classes** | `ConversationBufferMemory` (raw history)
`ConversationSummaryMemory` (LLM-summarized \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m history)
`ChatMessageHistory` (storage backend) | `ChatEngine` with modes:
`context` (standard \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m RAG)
`condense_question` (rewrites queries based on history) | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | **Context Strategy** | **Chains**: Sequential passing of context between steps.
**Agents**: Dynamic tool \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m selection to populate context. | **Indices**: `VectorStoreIndex` (semantic search), `ListIndex` (sequential), \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m `TreeIndex` (hierarchical summaries). | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m | **Optimization** | `Context Compression`: Removing irrelevant tokens from retrieved docs. | `Response \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Synthesizers`:
`compact` (stuffs max chunks into prompt)
`tree_summarize` (hierarchical reduction) | \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### 4. Handling Context Limits vs. Long-Context Models \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m A critical architectural decision is whether to use RAG or rely on massive context windows (e.g., Gemini 1.5 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Pro's 2M tokens). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **RAG + Small Context**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Pros**: Significantly cheaper (fewer tokens processed), lower latency, and often higher accuracy for \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m specific fact retrieval. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Cons**: Can miss \"global\" insights that require reading the entire corpus at once. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Long-Context Models (LCW)**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Pros**: Can ingest entire books or codebases, enabling \"whole-document\" reasoning. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Cons**: \"Lost in the Middle\" phenomenon where models forget information in the middle of a massive \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Trade-off**: Benchmarks show that RAG systems often outperform LCW models in \"Needle in a Haystack\" \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m tests for large corpora (2M+ tokens) because they filter noise before the model sees it [3]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### 5. Specific Patterns \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Needle in a Haystack Optimization**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **The Problem**: LLMs struggle to find a specific fact (\"needle\") buried in a large amount of \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m irrelevant text (\"haystack\"). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Optimization**: Place critical instructions or facts at the very **beginning** or very **end** of the \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context window, as models pay most attention to these areas (Serial Position Effect). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Stuffing vs. Context Selection**: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Stuffing**: Naively filling the window with as much data as possible. This often leads to \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m *context distraction* and higher costs [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * **Context Selection**: Intentionally curating what enters the window. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * *Techniques*: Using **Node Postprocessors** (LlamaIndex) to filter retrieved nodes by keywords or \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m similarity score before they reach the LLM [4]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m * *Query Transformations*: Rewriting a user's vague query into a precise search term to select better \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context (e.g., LangChain's `HyDE` or LlamaIndex's `Multi-step query decomposition`) [4]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m ### Sources \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m [2] Context Engineering Strategies (n8n): \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/ \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m [3] RAG vs Long Context Window Study: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m [4] LangChain & LlamaIndex Comparison: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m in-large-2b5f3fabc360 \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_todos                                                                                       │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"todos\": [                                                                                                    │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Save research request to /research_request.md\"                                                │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Research context engineering approaches for AI agents using a sub-agent\"                      │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"in_progress\",                                                                                  │\n",
       "│       \"content\": \"Synthesize findings and write final report to /final_report.md\"                               │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"pending\",                                                                                      │\n",
       "│       \"content\": \"Verify report against original request\"                                                       │\n",
       "│     }                                                                                                           │\n",
       "│   ]                                                                                                             │\n",
       "│ }                                                                                                               │\n",
       "│    ID: ea8aeaae-7e6a-4f8b-beba-1f252efabd24                                                                     │\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_file                                                                                        │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"file_path\": \"/final_report.md\",                                                                              │\n",
       "│   \"content\": \"# Context Engineering Approaches for AI Agents\\n\\n## 1. Introduction\\nContext Engineering is the  │\n",
       "│ systematic practice of designing, managing, and optimizing the information provided to AI models to ensure      │\n",
       "│ accurate, consistent, and context-aware behavior. While often confused with prompt engineering (which focuses   │\n",
       "│ on static instructions), context engineering treats the AI's context window as a dynamic, finite resource that  │\n",
       "│ must be architected and managed throughout the lifecycle of an interaction. It is a critical discipline for     │\n",
       "│ building robust AI agents that can maintain state and relevance over long, complex tasks.\\n\\n## 2. Core         │\n",
       "│ Concepts\\nTo build effective agents, developers must manage several fundamental aspects of the Large Language   │\n",
       "│ Model (LLM) environment:\\n\\n*   **Context Window**: The hard limit on the amount of text (tokens) an LLM can    │\n",
       "│ process at once (e.g., 8k, 128k, or 1M tokens). Effective engineering ensures this space is used                │\n",
       "│ efficiently.\\n*   **State Management**: Tracking the \\\"world state\\\" across multi-turn conversations. This      │\n",
       "│ includes user intent, task progress, and tool outputs.\\n*   **Context Pollution**: The phenomenon where         │\n",
       "│ irrelevant or contradictory information in the context degrades model performance, leading to hallucinations or │\n",
       "│ confusion [2].\\n*   **Serial Position Effect**: The tendency of LLMs to pay more attention to information at    │\n",
       "│ the very beginning and very end of the context window, often ignoring details buried in the middle (\\\"Lost in   │\n",
       "│ the Middle\\\") [3].\\n\\n## 3. Key Approaches and Techniques\\n\\n### Dynamic Context Injection & Retrieval          │\n",
       "│ (RAG)\\nRather than loading all available data at once, agents \\\"pull\\\" information only when necessary.\\n*      │\n",
       "│ **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time data only     │\n",
       "│ after identifying a specific knowledge gap [2].\\n*   **RAG (Retrieval-Augmented Generation)**: The standard for │\n",
       "│ large knowledge bases. Instead of \\\"stuffing\\\" a document, the system retrieves only the top-k most relevant    │\n",
       "│ chunks from a vector store to populate the context [2][3].\\n\\n### Memory Management\\nAgents require             │\n",
       "│ sophisticated memory systems to maintain continuity without overflowing context limits:\\n*   **Short-term       │\n",
       "│ Memory**: A sliding window of the most recent interaction turns (e.g., last 10 messages).\\n*   **Long-term      │\n",
       "│ Memory**: Key facts (e.g., user preferences) are persisted in external databases (Postgres, Vector Stores) and  │\n",
       "│ retrieved only when relevant to the current query [2].\\n*   **Pruning and Trimming**: Algorithms that           │\n",
       "│ selectively remove older, less relevant tokens while preserving the semantic core of the conversation           │\n",
       "│ history.\\n\\n### Summarization and Compression\\n*   **Tool Output Summarization**: Raw outputs from tools (like  │\n",
       "│ HTML from a web scraper) are often too verbose. An intermediate step or \\\"summarizer agent\\\" distills this data │\n",
       "│ into key points before passing it to the main reasoning agent [2].\\n*   **Context Compression**: Removing stop  │\n",
       "│ words or irrelevant sections from retrieved documents to maximize information density per token.\\n\\n### Prompt  │\n",
       "│ Structuring\\nUsing strict formats within the context window helps the model parse information correctly:\\n*     │\n",
       "│ **Separation of Concerns**: Clearly formatting \\\"System Instructions,\\\" \\\"User Input,\\\" and \\\"Tool Outputs\\\"    │\n",
       "│ (often using XML tags or JSON schemas) prevents the model from confusing data with instructions.\\n\\n## 4.       │\n",
       "│ Architectural Frameworks\\nTwo primary frameworks dominate the implementation of these                           │\n",
       "│ strategies:\\n\\n**LangChain**\\n*   **Focus**: Flexible orchestration and agent workflows.\\n*   **Approach**:     │\n",
       "│ Uses \\\"Chains\\\" to pass context sequentially between steps and \\\"Memory\\\" classes (e.g.,                        │\n",
       "│ `ConversationSummaryMemory`) to automatically manage history compression and persistence [4].\\n*                │\n",
       "│ **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to its own      │\n",
       "│ context.\\n\\n**LlamaIndex**\\n*   **Focus**: Data indexing and advanced retrieval.\\n*   **Approach**: specialized │\n",
       "│ in structuring data for LLMs using indices (`VectorStoreIndex`, `TreeIndex`).\\n*   **Strategy**: Uses           │\n",
       "│ \\\"Response Synthesizers\\\" to optimize how retrieved context is presented to the LLM (e.g., `compact` mode to    │\n",
       "│ stuff chunks or `tree_summarize` for hierarchical reduction) [4].\\n\\n## 5. Strategic Patterns\\n\\n### RAG vs.    │\n",
       "│ Long-Context Windows (LCW)\\n*   **RAG**: Preferred for cost efficiency and low latency. It filters noise before │\n",
       "│ the model sees it, often resulting in higher accuracy for \\\"Needle in a Haystack\\\" retrieval tasks [3].\\n*      │\n",
       "│ **Long-Context Models (e.g., Gemini 1.5)**: Best for \\\"whole-document\\\" reasoning where global context is       │\n",
       "│ required. However, they are more expensive and prone to the \\\"Lost in the Middle\\\" effect.\\n\\n### Context       │\n",
       "│ Selection vs. Stuffing\\n*   **Context Stuffing**: Naively filling the window with all available data. This is   │\n",
       "│ generally an anti-pattern as it increases costs and distracts the model.\\n*   **Context Selection**: The        │\n",
       "│ proactive curation of inputs. Techniques include \\\"Node Postprocessors\\\" (filtering results by similarity       │\n",
       "│ score) and \\\"Query Transformations\\\" (rewriting vague user queries to fetch better context) [4].\\n\\n## 6.       │\n",
       "│ Conclusion\\nContext engineering is the bridge between raw data and intelligent agent behavior. By moving from   │\n",
       "│ naive context stuffing to sophisticated dynamic injection, memory management, and structured retrieval,         │\n",
       "│ developers can build agents that are both powerful and cost-effective. The choice between RAG architectures and │\n",
       "│ long-context models remains a key trade-off, with hybrid approaches often providing the best balance of         │\n",
       "│ accuracy and performance.\\n\\n### Sources\\n[1] Context Engineering Guide:                                        │\n",
       "│ https://www.promptingguide.ai/agents/context-engineering\\n[2] Context Engineering Strategies (n8n):             │\n",
       "│ https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/\\n[3] RAG vs Long      │\n",
       "│ Context Window Study: https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs\\n[4]  │\n",
       "│ LangChain & LlamaIndex Comparison:                                                                              │\n",
       "│ https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- │\n",
       "│ in-large-2b5f3fabc366\"                                                                                          │\n",
       "│ }                                                                                                               │\n",
       "│    ID: 589357a0-fbce-4941-9090-8f91d12880cb                                                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_todos \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"todos\": [ \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Save research request to /research_request.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Research context engineering approaches for AI agents using a sub-agent\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"in_progress\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Synthesize findings and write final report to /final_report.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"pending\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Verify report against original request\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ] \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: ea8aeaae-7e6a-4f8b-beba-1f252efabd24 \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_file \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"file_path\": \"/final_report.md\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"# Context Engineering Approaches for AI Agents\\n\\n## 1. Introduction\\nContext Engineering is the \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m systematic practice of designing, managing, and optimizing the information provided to AI models to ensure \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m accurate, consistent, and context-aware behavior. While often confused with prompt engineering (which focuses \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m on static instructions), context engineering treats the AI's context window as a dynamic, finite resource that \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m must be architected and managed throughout the lifecycle of an interaction. It is a critical discipline for \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m building robust AI agents that can maintain state and relevance over long, complex tasks.\\n\\n## 2. Core \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Concepts\\nTo build effective agents, developers must manage several fundamental aspects of the Large Language \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Model (LLM) environment:\\n\\n* **Context Window**: The hard limit on the amount of text (tokens) an LLM can \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m process at once (e.g., 8k, 128k, or 1M tokens). Effective engineering ensures this space is used \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m efficiently.\\n* **State Management**: Tracking the \\\"world state\\\" across multi-turn conversations. This \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m includes user intent, task progress, and tool outputs.\\n* **Context Pollution**: The phenomenon where \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m irrelevant or contradictory information in the context degrades model performance, leading to hallucinations or \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m confusion [2].\\n* **Serial Position Effect**: The tendency of LLMs to pay more attention to information at \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m the very beginning and very end of the context window, often ignoring details buried in the middle (\\\"Lost in \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m the Middle\\\") [3].\\n\\n## 3. Key Approaches and Techniques\\n\\n### Dynamic Context Injection & Retrieval \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m (RAG)\\nRather than loading all available data at once, agents \\\"pull\\\" information only when necessary.\\n* \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time data only \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m after identifying a specific knowledge gap [2].\\n* **RAG (Retrieval-Augmented Generation)**: The standard for \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m large knowledge bases. Instead of \\\"stuffing\\\" a document, the system retrieves only the top-k most relevant \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m chunks from a vector store to populate the context [2][3].\\n\\n### Memory Management\\nAgents require \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m sophisticated memory systems to maintain continuity without overflowing context limits:\\n* **Short-term \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Memory**: A sliding window of the most recent interaction turns (e.g., last 10 messages).\\n* **Long-term \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Memory**: Key facts (e.g., user preferences) are persisted in external databases (Postgres, Vector Stores) and \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m retrieved only when relevant to the current query [2].\\n* **Pruning and Trimming**: Algorithms that \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m selectively remove older, less relevant tokens while preserving the semantic core of the conversation \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m history.\\n\\n### Summarization and Compression\\n* **Tool Output Summarization**: Raw outputs from tools (like \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m HTML from a web scraper) are often too verbose. An intermediate step or \\\"summarizer agent\\\" distills this data \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m into key points before passing it to the main reasoning agent [2].\\n* **Context Compression**: Removing stop \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m words or irrelevant sections from retrieved documents to maximize information density per token.\\n\\n### Prompt \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Structuring\\nUsing strict formats within the context window helps the model parse information correctly:\\n* \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Separation of Concerns**: Clearly formatting \\\"System Instructions,\\\" \\\"User Input,\\\" and \\\"Tool Outputs\\\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m (often using XML tags or JSON schemas) prevents the model from confusing data with instructions.\\n\\n## 4. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Architectural Frameworks\\nTwo primary frameworks dominate the implementation of these \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m strategies:\\n\\n**LangChain**\\n* **Focus**: Flexible orchestration and agent workflows.\\n* **Approach**: \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Uses \\\"Chains\\\" to pass context sequentially between steps and \\\"Memory\\\" classes (e.g., \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m `ConversationSummaryMemory`) to automatically manage history compression and persistence [4].\\n* \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to its own \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m context.\\n\\n**LlamaIndex**\\n* **Focus**: Data indexing and advanced retrieval.\\n* **Approach**: specialized \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m in structuring data for LLMs using indices (`VectorStoreIndex`, `TreeIndex`).\\n* **Strategy**: Uses \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \\\"Response Synthesizers\\\" to optimize how retrieved context is presented to the LLM (e.g., `compact` mode to \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m stuff chunks or `tree_summarize` for hierarchical reduction) [4].\\n\\n## 5. Strategic Patterns\\n\\n### RAG vs. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Long-Context Windows (LCW)\\n* **RAG**: Preferred for cost efficiency and low latency. It filters noise before \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m the model sees it, often resulting in higher accuracy for \\\"Needle in a Haystack\\\" retrieval tasks [3].\\n* \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Long-Context Models (e.g., Gemini 1.5)**: Best for \\\"whole-document\\\" reasoning where global context is \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m required. However, they are more expensive and prone to the \\\"Lost in the Middle\\\" effect.\\n\\n### Context \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Selection vs. Stuffing\\n* **Context Stuffing**: Naively filling the window with all available data. This is \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m generally an anti-pattern as it increases costs and distracts the model.\\n* **Context Selection**: The \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m proactive curation of inputs. Techniques include \\\"Node Postprocessors\\\" (filtering results by similarity \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m score) and \\\"Query Transformations\\\" (rewriting vague user queries to fetch better context) [4].\\n\\n## 6. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Conclusion\\nContext engineering is the bridge between raw data and intelligent agent behavior. By moving from \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m naive context stuffing to sophisticated dynamic injection, memory management, and structured retrieval, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m developers can build agents that are both powerful and cost-effective. The choice between RAG architectures and \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m long-context models remains a key trade-off, with hybrid approaches often providing the best balance of \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m accuracy and performance.\\n\\n### Sources\\n[1] Context Engineering Guide: \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m https://www.promptingguide.ai/agents/context-engineering\\n[2] Context Engineering Strategies (n8n): \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/\\n[3] RAG vs Long \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Context Window Study: https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs\\n[4] \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m LangChain & LlamaIndex Comparison: \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m in-large-2b5f3fabc366\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: 589357a0-fbce-4941-9090-8f91d12880cb \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'},      │\n",
       "│ {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'completed'},  │\n",
       "│ {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'in_progress'},         │\n",
       "│ {'content': 'Verify report against original request', 'status': 'pending'}]                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'in_progress'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Verify report against original request', 'status': 'pending'}] \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated file /final_report.md                                                                                   │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated file /final_report.md \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: read_file                                                                                         │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"file_path\": \"/research_request.md\"                                                                           │\n",
       "│ }                                                                                                               │\n",
       "│    ID: dacd96a0-c106-462a-a710-6fc389064df5                                                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: read_file \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"file_path\": \"/research_request.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: dacd96a0-c106-462a-a710-6fc389064df5 \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│      1  research context engineering approaches used to build AI agents                                         │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m 1 research context engineering approaches used to build AI agents \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: write_todos                                                                                       │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"todos\": [                                                                                                    │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Save research request to /research_request.md\"                                                │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Research context engineering approaches for AI agents using a sub-agent\"                      │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Synthesize findings and write final report to /final_report.md\"                               │\n",
       "│     },                                                                                                          │\n",
       "│     {                                                                                                           │\n",
       "│       \"status\": \"completed\",                                                                                    │\n",
       "│       \"content\": \"Verify report against original request\"                                                       │\n",
       "│     }                                                                                                           │\n",
       "│   ]                                                                                                             │\n",
       "│ }                                                                                                               │\n",
       "│    ID: 4c420284-baa7-4a5b-90ab-e156b502c709                                                                     │\n",
       "│                                                                                                                 │\n",
       "│ 🔧 Tool Call: read_file                                                                                         │\n",
       "│    Args: {                                                                                                      │\n",
       "│   \"file_path\": \"/final_report.md\"                                                                               │\n",
       "│ }                                                                                                               │\n",
       "│    ID: 2d15c15f-589e-4c3d-98b0-03116707c143                                                                     │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: write_todos \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"todos\": [ \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Save research request to /research_request.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Research context engineering approaches for AI agents using a sub-agent\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Synthesize findings and write final report to /final_report.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m }, \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"status\": \"completed\", \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"content\": \"Verify report against original request\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ] \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: 4c420284-baa7-4a5b-90ab-e156b502c709 \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 🔧 Tool Call: read_file \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Args: { \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \"file_path\": \"/final_report.md\" \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m } \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ID: 2d15c15f-589e-4c3d-98b0-03116707c143 \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│ Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'},      │\n",
       "│ {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'completed'},  │\n",
       "│ {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'completed'},           │\n",
       "│ {'content': 'Verify report against original request', 'status': 'completed'}]                                   │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m Updated todo list to [{'content': 'Save research request to /research_request.md', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Research context engineering approaches for AI agents using a sub-agent', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Synthesize findings and write final report to /final_report.md', 'status': 'completed'}, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m {'content': 'Verify report against original request', 'status': 'completed'}] \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭──────────────────────────────────────────────── 🔧 Tool Output ─────────────────────────────────────────────────╮\n",
       "│      1  # Context Engineering Approaches for AI Agents                                                          │\n",
       "│      2                                                                                                          │\n",
       "│      3  ## 1. Introduction                                                                                      │\n",
       "│      4  Context Engineering is the systematic practice of designing, managing, and optimizing the information   │\n",
       "│ provided to AI models to ensure accurate, consistent, and context-aware behavior. While often confused with     │\n",
       "│ prompt engineering (which focuses on static instructions), context engineering treats the AI's context window   │\n",
       "│ as a dynamic, finite resource that must be architected and managed throughout the lifecycle of an interaction.  │\n",
       "│ It is a critical discipline for building robust AI agents that can maintain state and relevance over long,      │\n",
       "│ complex tasks.                                                                                                  │\n",
       "│      5                                                                                                          │\n",
       "│      6  ## 2. Core Concepts                                                                                     │\n",
       "│      7  To build effective agents, developers must manage several fundamental aspects of the Large Language     │\n",
       "│ Model (LLM) environment:                                                                                        │\n",
       "│      8                                                                                                          │\n",
       "│      9  *   **Context Window**: The hard limit on the amount of text (tokens) an LLM can process at once (e.g., │\n",
       "│ 8k, 128k, or 1M tokens). Effective engineering ensures this space is used efficiently.                          │\n",
       "│     10  *   **State Management**: Tracking the \"world state\" across multi-turn conversations. This includes     │\n",
       "│ user intent, task progress, and tool outputs.                                                                   │\n",
       "│     11  *   **Context Pollution**: The phenomenon where irrelevant or contradictory information in the context  │\n",
       "│ degrades model performance, leading to hallucinations or confusion [2].                                         │\n",
       "│     12  *   **Serial Position Effect**: The tendency of LLMs to pay more attention to information at the very   │\n",
       "│ beginning and very end of the context window, often ignoring details buried in the middle (\"Lost in the         │\n",
       "│ Middle\") [3].                                                                                                   │\n",
       "│     13                                                                                                          │\n",
       "│     14  ## 3. Key Approaches and Techniques                                                                     │\n",
       "│     15                                                                                                          │\n",
       "│     16  ### Dynamic Context Injection & Retrieval (RAG)                                                         │\n",
       "│     17  Rather than loading all available data at once, agents \"pull\" information only when necessary.          │\n",
       "│     18  *   **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time   │\n",
       "│ data only after identifying a specific knowledge gap [2].                                                       │\n",
       "│     19  *   **RAG (Retrieval-Augmented Generation)**: The standard for large knowledge bases. Instead of        │\n",
       "│ \"stuffing\" a document, the system retrieves only the top-k most relevant chunks from a vector store to populate │\n",
       "│ the context [2][3].                                                                                             │\n",
       "│     20                                                                                                          │\n",
       "│     21  ### Memory Management                                                                                   │\n",
       "│     22  Agents require sophisticated memory systems to maintain continuity without overflowing context limits:  │\n",
       "│     23  *   **Short-term Memory**: A sliding window of the most recent interaction turns (e.g., last 10         │\n",
       "│ messages).                                                                                                      │\n",
       "│     24  *   **Long-term Memory**: Key facts (e.g., user preferences) are persisted in external databases        │\n",
       "│ (Postgres, Vector Stores) and retrieved only when relevant to the current query [2].                            │\n",
       "│     25  *   **Pruning and Trimming**: Algorithms that selectively remove older, less relevant tokens while      │\n",
       "│ preserving the semantic core of the conversation history.                                                       │\n",
       "│     26                                                                                                          │\n",
       "│     27  ### Summarization and Compression                                                                       │\n",
       "│     28  *   **Tool Output Summarization**: Raw outputs from tools (like HTML from a web scraper) are often too  │\n",
       "│ verbose. An intermediate step or \"summarizer agent\" distills this data into key points before passing it to the │\n",
       "│ main reasoning agent [2].                                                                                       │\n",
       "│     29  *   **Context Compression**: Removing stop words or irrelevant sections from retrieved documents to     │\n",
       "│ maximize information density per token.                                                                         │\n",
       "│     30                                                                                                          │\n",
       "│     31  ### Prompt Structuring                                                                                  │\n",
       "│     32  Using strict formats within the context window helps the model parse information correctly:             │\n",
       "│     33  *   **Separation of Concerns**: Clearly formatting \"System Instructions,\" \"User Input,\" and \"Tool       │\n",
       "│ Outputs\" (often using XML tags or JSON schemas) prevents the model from confusing data with instructions.       │\n",
       "│     34                                                                                                          │\n",
       "│     35  ## 4. Architectural Frameworks                                                                          │\n",
       "│     36  Two primary frameworks dominate the implementation of these strategies:                                 │\n",
       "│     37                                                                                                          │\n",
       "│     38  **LangChain**                                                                                           │\n",
       "│     39  *   **Focus**: Flexible orchestration and agent workflows.                                              │\n",
       "│     40  *   **Approach**: Uses \"Chains\" to pass context sequentially between steps and \"Memory\" classes (e.g.,  │\n",
       "│ `ConversationSummaryMemory`) to automatically manage history compression and persistence [4].                   │\n",
       "│     41  *   **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to  │\n",
       "│ its own context.                                                                                                │\n",
       "│     42                                                                                                          │\n",
       "│     43  **LlamaIndex**                                                                                          │\n",
       "│     44  *   **Focus**: Data indexing and advanced retrieval.                                                    │\n",
       "│     45  *   **Approach**: specialized in structuring data for LLMs using indices (`VectorStoreIndex`,           │\n",
       "│ `TreeIndex`).                                                                                                   │\n",
       "│     46  *   **Strategy**: Uses \"Response Synthesizers\" to optimize how retrieved context is presented to the    │\n",
       "│ LLM (e.g., `compact` mode to stuff chunks or `tree_summarize` for hierarchical reduction) [4].                  │\n",
       "│     47                                                                                                          │\n",
       "│     48  ## 5. Strategic Patterns                                                                                │\n",
       "│     49                                                                                                          │\n",
       "│     50  ### RAG vs. Long-Context Windows (LCW)                                                                  │\n",
       "│     51  *   **RAG**: Preferred for cost efficiency and low latency. It filters noise before the model sees it,  │\n",
       "│ often resulting in higher accuracy for \"Needle in a Haystack\" retrieval tasks [3].                              │\n",
       "│     52  *   **Long-Context Models (e.g., Gemini 1.5)**: Best for \"whole-document\" reasoning where global        │\n",
       "│ context is required. However, they are more expensive and prone to the \"Lost in the Middle\" effect.             │\n",
       "│     53                                                                                                          │\n",
       "│     54  ### Context Selection vs. Stuffing                                                                      │\n",
       "│     55  *   **Context Stuffing**: Naively filling the window with all available data. This is generally an      │\n",
       "│ anti-pattern as it increases costs and distracts the model.                                                     │\n",
       "│     56  *   **Context Selection**: The proactive curation of inputs. Techniques include \"Node Postprocessors\"   │\n",
       "│ (filtering results by similarity score) and \"Query Transformations\" (rewriting vague user queries to fetch      │\n",
       "│ better context) [4].                                                                                            │\n",
       "│     57                                                                                                          │\n",
       "│     58  ## 6. Conclusion                                                                                        │\n",
       "│     59  Context engineering is the bridge between raw data and intelligent agent behavior. By moving from naive │\n",
       "│ context stuffing to sophisticated dynamic injection, memory management, and structured retrieval, developers    │\n",
       "│ can build agents that are both powerful and cost-effective. The choice between RAG architectures and            │\n",
       "│ long-context models remains a key trade-off, with hybrid approaches often providing the best balance of         │\n",
       "│ accuracy and performance.                                                                                       │\n",
       "│     60                                                                                                          │\n",
       "│     61  ### Sources                                                                                             │\n",
       "│     62  [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering                 │\n",
       "│     63  [2] Context Engineering Strategies (n8n):                                                               │\n",
       "│ https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/                       │\n",
       "│     64  [3] RAG vs Long Context Window Study:                                                                   │\n",
       "│ https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs                             │\n",
       "│     65  [4] LangChain & LlamaIndex Comparison:                                                                  │\n",
       "│ https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- │\n",
       "│ in-large-2b5f3fabc366                                                                                           │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[33m╭─\u001b[0m\u001b[33m───────────────────────────────────────────────\u001b[0m\u001b[33m 🔧 Tool Output \u001b[0m\u001b[33m────────────────────────────────────────────────\u001b[0m\u001b[33m─╮\u001b[0m\n", "\u001b[33m│\u001b[0m 1 # Context Engineering Approaches for AI Agents \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 2 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 3 ## 1. Introduction \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 4 Context Engineering is the systematic practice of designing, managing, and optimizing the information \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m provided to AI models to ensure accurate, consistent, and context-aware behavior. While often confused with \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m prompt engineering (which focuses on static instructions), context engineering treats the AI's context window \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m as a dynamic, finite resource that must be architected and managed throughout the lifecycle of an interaction. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m It is a critical discipline for building robust AI agents that can maintain state and relevance over long, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m complex tasks. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 5 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 6 ## 2. Core Concepts \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 7 To build effective agents, developers must manage several fundamental aspects of the Large Language \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Model (LLM) environment: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 8 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 9 * **Context Window**: The hard limit on the amount of text (tokens) an LLM can process at once (e.g., \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 8k, 128k, or 1M tokens). Effective engineering ensures this space is used efficiently. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 10 * **State Management**: Tracking the \"world state\" across multi-turn conversations. This includes \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m user intent, task progress, and tool outputs. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 11 * **Context Pollution**: The phenomenon where irrelevant or contradictory information in the context \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m degrades model performance, leading to hallucinations or confusion [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 12 * **Serial Position Effect**: The tendency of LLMs to pay more attention to information at the very \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m beginning and very end of the context window, often ignoring details buried in the middle (\"Lost in the \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Middle\") [3]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 13 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 14 ## 3. Key Approaches and Techniques \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 15 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 16 ### Dynamic Context Injection & Retrieval (RAG) \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 17 Rather than loading all available data at once, agents \"pull\" information only when necessary. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 18 * **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m data only after identifying a specific knowledge gap [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 19 * **RAG (Retrieval-Augmented Generation)**: The standard for large knowledge bases. Instead of \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m \"stuffing\" a document, the system retrieves only the top-k most relevant chunks from a vector store to populate \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m the context [2][3]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 20 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 21 ### Memory Management \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 22 Agents require sophisticated memory systems to maintain continuity without overflowing context limits: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 23 * **Short-term Memory**: A sliding window of the most recent interaction turns (e.g., last 10 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m messages). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 24 * **Long-term Memory**: Key facts (e.g., user preferences) are persisted in external databases \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m (Postgres, Vector Stores) and retrieved only when relevant to the current query [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 25 * **Pruning and Trimming**: Algorithms that selectively remove older, less relevant tokens while \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m preserving the semantic core of the conversation history. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 26 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 27 ### Summarization and Compression \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 28 * **Tool Output Summarization**: Raw outputs from tools (like HTML from a web scraper) are often too \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m verbose. An intermediate step or \"summarizer agent\" distills this data into key points before passing it to the \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m main reasoning agent [2]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 29 * **Context Compression**: Removing stop words or irrelevant sections from retrieved documents to \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m maximize information density per token. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 30 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 31 ### Prompt Structuring \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 32 Using strict formats within the context window helps the model parse information correctly: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 33 * **Separation of Concerns**: Clearly formatting \"System Instructions,\" \"User Input,\" and \"Tool \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m Outputs\" (often using XML tags or JSON schemas) prevents the model from confusing data with instructions. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 34 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 35 ## 4. Architectural Frameworks \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 36 Two primary frameworks dominate the implementation of these strategies: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 37 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 38 **LangChain** \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 39 * **Focus**: Flexible orchestration and agent workflows. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 40 * **Approach**: Uses \"Chains\" to pass context sequentially between steps and \"Memory\" classes (e.g., \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m `ConversationSummaryMemory`) to automatically manage history compression and persistence [4]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 41 * **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m its own context. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 42 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 43 **LlamaIndex** \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 44 * **Focus**: Data indexing and advanced retrieval. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 45 * **Approach**: specialized in structuring data for LLMs using indices (`VectorStoreIndex`, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m `TreeIndex`). \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 46 * **Strategy**: Uses \"Response Synthesizers\" to optimize how retrieved context is presented to the \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m LLM (e.g., `compact` mode to stuff chunks or `tree_summarize` for hierarchical reduction) [4]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 47 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 48 ## 5. Strategic Patterns \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 49 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 50 ### RAG vs. Long-Context Windows (LCW) \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 51 * **RAG**: Preferred for cost efficiency and low latency. It filters noise before the model sees it, \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m often resulting in higher accuracy for \"Needle in a Haystack\" retrieval tasks [3]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 52 * **Long-Context Models (e.g., Gemini 1.5)**: Best for \"whole-document\" reasoning where global \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context is required. However, they are more expensive and prone to the \"Lost in the Middle\" effect. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 53 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 54 ### Context Selection vs. Stuffing \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 55 * **Context Stuffing**: Naively filling the window with all available data. This is generally an \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m anti-pattern as it increases costs and distracts the model. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 56 * **Context Selection**: The proactive curation of inputs. Techniques include \"Node Postprocessors\" \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m (filtering results by similarity score) and \"Query Transformations\" (rewriting vague user queries to fetch \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m better context) [4]. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 57 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 58 ## 6. Conclusion \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 59 Context engineering is the bridge between raw data and intelligent agent behavior. By moving from naive \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m context stuffing to sophisticated dynamic injection, memory management, and structured retrieval, developers \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m can build agents that are both powerful and cost-effective. The choice between RAG architectures and \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m long-context models remains a key trade-off, with hybrid approaches often providing the best balance of \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m accuracy and performance. \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 60 \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 61 ### Sources \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 62 [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 63 [2] Context Engineering Strategies (n8n): \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/ \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 64 [3] RAG vs Long Context Window Study: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m 65 [4] LangChain & LlamaIndex Comparison: \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperability- \u001b[33m│\u001b[0m\n", "\u001b[33m│\u001b[0m in-large-2b5f3fabc366 \u001b[33m│\u001b[0m\n", "\u001b[33m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
╭───────────────────────────────────────────────────── 📝 AI ─────────────────────────────────────────────────────╮\n",
       "│ I have completed the research on context engineering approaches for AI agents. The final report details         │\n",
       "│ definitions, core concepts, key techniques (RAG, Memory, Summarization), and architectural frameworks           │\n",
       "│ (LangChain vs LlamaIndex).                                                                                      │\n",
       "│                                                                                                                 │\n",
       "│ ### **Research Findings Summary**                                                                               │\n",
       "│ Context engineering is the discipline of managing the finite \"context window\" of an AI model to ensure it has   │\n",
       "│ the right information at the right time. It is distinct from prompt engineering and focuses on dynamic state    │\n",
       "│ management.                                                                                                     │\n",
       "│                                                                                                                 │\n",
       "│ **Key Approaches:**                                                                                             │\n",
       "│ 1.  **Dynamic Context Injection (RAG)**: Instead of loading all data, agents \"pull\" specific information from   │\n",
       "│ vector databases or APIs only when needed to answer a query.                                                    │\n",
       "│ 2.  **Memory Management**: Using sliding windows (short-term memory) and external databases (long-term memory)  │\n",
       "│ to maintain conversation history without exceeding token limits.                                                │\n",
       "│ 3.  **Summarization**: Compressing verbose tool outputs (like web scrape results) before passing them to the    │\n",
       "│ reasoning agent to save tokens.                                                                                 │\n",
       "│ 4.  **Architectural Frameworks**:                                                                               │\n",
       "│     *   **LangChain**: Focuses on \"Chains\" and dynamic tool selection.                                          │\n",
       "│     *   **LlamaIndex**: Focuses on data indexing and advanced retrieval strategies.                             │\n",
       "│                                                                                                                 │\n",
       "│ **Strategic Trade-offs:**                                                                                       │\n",
       "│ *   **RAG vs. Long-Context**: RAG is often more cost-effective and accurate for finding specific facts          │\n",
       "│ (\"needles\") in large datasets, while massive context windows (e.g., 1M+ tokens) are better for holistic         │\n",
       "│ analysis but prone to the \"Lost in the Middle\" phenomenon.                                                      │\n",
       "│ *   **Selection vs. Stuffing**: \"Context Stuffing\" (filling the window blindly) is an anti-pattern. \"Context    │\n",
       "│ Selection\" (curating inputs via filtering and query transformation) leads to better performance.                │\n",
       "│                                                                                                                 │\n",
       "│ The full detailed report can be found in `/final_report.md`.                                                    │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[37m╭─\u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m 📝 AI \u001b[0m\u001b[37m────────────────────────────────────────────────────\u001b[0m\u001b[37m─╮\u001b[0m\n", "\u001b[37m│\u001b[0m I have completed the research on context engineering approaches for AI agents. The final report details \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m definitions, core concepts, key techniques (RAG, Memory, Summarization), and architectural frameworks \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m (LangChain vs LlamaIndex). \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m ### **Research Findings Summary** \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Context engineering is the discipline of managing the finite \"context window\" of an AI model to ensure it has \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m the right information at the right time. It is distinct from prompt engineering and focuses on dynamic state \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m management. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Key Approaches:** \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 1. **Dynamic Context Injection (RAG)**: Instead of loading all data, agents \"pull\" specific information from \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m vector databases or APIs only when needed to answer a query. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 2. **Memory Management**: Using sliding windows (short-term memory) and external databases (long-term memory) \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m to maintain conversation history without exceeding token limits. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 3. **Summarization**: Compressing verbose tool outputs (like web scrape results) before passing them to the \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m reasoning agent to save tokens. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m 4. **Architectural Frameworks**: \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m * **LangChain**: Focuses on \"Chains\" and dynamic tool selection. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m * **LlamaIndex**: Focuses on data indexing and advanced retrieval strategies. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m **Strategic Trade-offs:** \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m * **RAG vs. Long-Context**: RAG is often more cost-effective and accurate for finding specific facts \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m (\"needles\") in large datasets, while massive context windows (e.g., 1M+ tokens) are better for holistic \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m analysis but prone to the \"Lost in the Middle\" phenomenon. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m * **Selection vs. Stuffing**: \"Context Stuffing\" (filling the window blindly) is an anti-pattern. \"Context \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m Selection\" (curating inputs via filtering and query transformation) leads to better performance. \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m \u001b[37m│\u001b[0m\n", "\u001b[37m│\u001b[0m The full detailed report can be found in `/final_report.md`. \u001b[37m│\u001b[0m\n", "\u001b[37m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "result = agent.invoke(\n", " {\n", " \"messages\": [\n", " {\n", " \"role\": \"user\",\n", " \"content\": \"research context engineering approaches used to build AI agents\",\n", " }\n", " ],\n", " }, \n", ")\n", "format_messages(result[\"messages\"])" ] }, { "cell_type": "code", "execution_count": 13, "id": "188b5ab5", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
╭──────────────────────────────────────────────────── Prompt ─────────────────────────────────────────────────────╮\n",
       "│                                                                                                                 │\n",
       "│  # Context Engineering Approaches for AI Agents                                                                 │\n",
       "│                                                                                                                 │\n",
       "│  ## 1. Introduction                                                                                             │\n",
       "│  Context Engineering is the systematic practice of designing, managing, and optimizing the information          │\n",
       "│  provided to AI models to ensure accurate, consistent, and context-aware behavior. While often confused with    │\n",
       "│  prompt engineering (which focuses on static instructions), context engineering treats the AI's context window  │\n",
       "│  as a dynamic, finite resource that must be architected and managed throughout the lifecycle of an              │\n",
       "│  interaction. It is a critical discipline for building robust AI agents that can maintain state and relevance   │\n",
       "│  over long, complex tasks.                                                                                      │\n",
       "│                                                                                                                 │\n",
       "│  ## 2. Core Concepts                                                                                            │\n",
       "│  To build effective agents, developers must manage several fundamental aspects of the Large Language Model      │\n",
       "│  (LLM) environment:                                                                                             │\n",
       "│                                                                                                                 │\n",
       "│  *   **Context Window**: The hard limit on the amount of text (tokens) an LLM can process at once (e.g., 8k,    │\n",
       "│  128k, or 1M tokens). Effective engineering ensures this space is used efficiently.                             │\n",
       "│  *   **State Management**: Tracking the \"world state\" across multi-turn conversations. This includes user       │\n",
       "│  intent, task progress, and tool outputs.                                                                       │\n",
       "│  *   **Context Pollution**: The phenomenon where irrelevant or contradictory information in the context         │\n",
       "│  degrades model performance, leading to hallucinations or confusion [2].                                        │\n",
       "│  *   **Serial Position Effect**: The tendency of LLMs to pay more attention to information at the very          │\n",
       "│  beginning and very end of the context window, often ignoring details buried in the middle (\"Lost in the        │\n",
       "│  Middle\") [3].                                                                                                  │\n",
       "│                                                                                                                 │\n",
       "│  ## 3. Key Approaches and Techniques                                                                            │\n",
       "│                                                                                                                 │\n",
       "│  ### Dynamic Context Injection & Retrieval (RAG)                                                                │\n",
       "│  Rather than loading all available data at once, agents \"pull\" information only when necessary.                 │\n",
       "│  *   **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time data     │\n",
       "│  only after identifying a specific knowledge gap [2].                                                           │\n",
       "│  *   **RAG (Retrieval-Augmented Generation)**: The standard for large knowledge bases. Instead of \"stuffing\" a  │\n",
       "│  document, the system retrieves only the top-k most relevant chunks from a vector store to populate the         │\n",
       "│  context [2][3].                                                                                                │\n",
       "│                                                                                                                 │\n",
       "│  ### Memory Management                                                                                          │\n",
       "│  Agents require sophisticated memory systems to maintain continuity without overflowing context limits:         │\n",
       "│  *   **Short-term Memory**: A sliding window of the most recent interaction turns (e.g., last 10 messages).     │\n",
       "│  *   **Long-term Memory**: Key facts (e.g., user preferences) are persisted in external databases (Postgres,    │\n",
       "│  Vector Stores) and retrieved only when relevant to the current query [2].                                      │\n",
       "│  *   **Pruning and Trimming**: Algorithms that selectively remove older, less relevant tokens while preserving  │\n",
       "│  the semantic core of the conversation history.                                                                 │\n",
       "│                                                                                                                 │\n",
       "│  ### Summarization and Compression                                                                              │\n",
       "│  *   **Tool Output Summarization**: Raw outputs from tools (like HTML from a web scraper) are often too         │\n",
       "│  verbose. An intermediate step or \"summarizer agent\" distills this data into key points before passing it to    │\n",
       "│  the main reasoning agent [2].                                                                                  │\n",
       "│  *   **Context Compression**: Removing stop words or irrelevant sections from retrieved documents to maximize   │\n",
       "│  information density per token.                                                                                 │\n",
       "│                                                                                                                 │\n",
       "│  ### Prompt Structuring                                                                                         │\n",
       "│  Using strict formats within the context window helps the model parse information correctly:                    │\n",
       "│  *   **Separation of Concerns**: Clearly formatting \"System Instructions,\" \"User Input,\" and \"Tool Outputs\"     │\n",
       "│  (often using XML tags or JSON schemas) prevents the model from confusing data with instructions.               │\n",
       "│                                                                                                                 │\n",
       "│  ## 4. Architectural Frameworks                                                                                 │\n",
       "│  Two primary frameworks dominate the implementation of these strategies:                                        │\n",
       "│                                                                                                                 │\n",
       "│  **LangChain**                                                                                                  │\n",
       "│  *   **Focus**: Flexible orchestration and agent workflows.                                                     │\n",
       "│  *   **Approach**: Uses \"Chains\" to pass context sequentially between steps and \"Memory\" classes (e.g.,         │\n",
       "│  `ConversationSummaryMemory`) to automatically manage history compression and persistence [4].                  │\n",
       "│  *   **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to its     │\n",
       "│  own context.                                                                                                   │\n",
       "│                                                                                                                 │\n",
       "│  **LlamaIndex**                                                                                                 │\n",
       "│  *   **Focus**: Data indexing and advanced retrieval.                                                           │\n",
       "│  *   **Approach**: specialized in structuring data for LLMs using indices (`VectorStoreIndex`, `TreeIndex`).    │\n",
       "│  *   **Strategy**: Uses \"Response Synthesizers\" to optimize how retrieved context is presented to the LLM       │\n",
       "│  (e.g., `compact` mode to stuff chunks or `tree_summarize` for hierarchical reduction) [4].                     │\n",
       "│                                                                                                                 │\n",
       "│  ## 5. Strategic Patterns                                                                                       │\n",
       "│                                                                                                                 │\n",
       "│  ### RAG vs. Long-Context Windows (LCW)                                                                         │\n",
       "│  *   **RAG**: Preferred for cost efficiency and low latency. It filters noise before the model sees it, often   │\n",
       "│  resulting in higher accuracy for \"Needle in a Haystack\" retrieval tasks [3].                                   │\n",
       "│  *   **Long-Context Models (e.g., Gemini 1.5)**: Best for \"whole-document\" reasoning where global context is    │\n",
       "│  required. However, they are more expensive and prone to the \"Lost in the Middle\" effect.                       │\n",
       "│                                                                                                                 │\n",
       "│  ### Context Selection vs. Stuffing                                                                             │\n",
       "│  *   **Context Stuffing**: Naively filling the window with all available data. This is generally an             │\n",
       "│  anti-pattern as it increases costs and distracts the model.                                                    │\n",
       "│  *   **Context Selection**: The proactive curation of inputs. Techniques include \"Node Postprocessors\"          │\n",
       "│  (filtering results by similarity score) and \"Query Transformations\" (rewriting vague user queries to fetch     │\n",
       "│  better context) [4].                                                                                           │\n",
       "│                                                                                                                 │\n",
       "│  ## 6. Conclusion                                                                                               │\n",
       "│  Context engineering is the bridge between raw data and intelligent agent behavior. By moving from naive        │\n",
       "│  context stuffing to sophisticated dynamic injection, memory management, and structured retrieval, developers   │\n",
       "│  can build agents that are both powerful and cost-effective. The choice between RAG architectures and           │\n",
       "│  long-context models remains a key trade-off, with hybrid approaches often providing the best balance of        │\n",
       "│  accuracy and performance.                                                                                      │\n",
       "│                                                                                                                 │\n",
       "│  ### Sources                                                                                                    │\n",
       "│  [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering                        │\n",
       "│  [2] Context Engineering Strategies (n8n):                                                                      │\n",
       "│  https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/                      │\n",
       "│  [3] RAG vs Long Context Window Study:                                                                          │\n",
       "│  https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs                            │\n",
       "│  [4] LangChain & LlamaIndex Comparison:                                                                         │\n",
       "│  https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperabilit  │\n",
       "│  y-in-large-2b5f3fabc366                                                                                        │\n",
       "│                                                                                                                 │\n",
       "╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\n",
       "
\n" ], "text/plain": [ "\u001b[34m╭─\u001b[0m\u001b[34m───────────────────────────────────────────────────\u001b[0m\u001b[34m \u001b[0m\u001b[1;32mPrompt\u001b[0m\u001b[34m \u001b[0m\u001b[34m────────────────────────────────────────────────────\u001b[0m\u001b[34m─╮\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m # Context Engineering Approaches for AI Agents \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 1. Introduction\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Context Engineering is the systematic practice of designing, managing, and optimizing the information \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m provided to AI models to ensure accurate, consistent, and context-aware behavior. While often confused with \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m prompt engineering (which focuses on static instructions), context engineering treats the AI's context window \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m as a dynamic, finite resource that must be architected and managed throughout the lifecycle of an \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m interaction. It is a critical discipline for building robust AI agents that can maintain state and relevance \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m over long, complex tasks. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 2. Core Concepts\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m To build effective agents, developers must manage several fundamental aspects of the Large Language Model \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m (LLM) environment: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Context Window**: The hard limit on the amount of text (tokens) an LLM can process at once (e.g., 8k, \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m 128k, or 1M tokens). Effective engineering ensures this space is used efficiently. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **State Management**: Tracking the \"world state\" across multi-turn conversations. This includes user \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m intent, task progress, and tool outputs. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Context Pollution**: The phenomenon where irrelevant or contradictory information in the context \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m degrades model performance, leading to hallucinations or confusion [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Serial Position Effect**: The tendency of LLMs to pay more attention to information at the very \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m beginning and very end of the context window, often ignoring details buried in the middle (\"Lost in the \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Middle\") [3]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 3. Key Approaches and Techniques\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Dynamic Context Injection & Retrieval (RAG)\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Rather than loading all available data at once, agents \"pull\" information only when necessary. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Dynamic Injection**: Agents use tools (like search APIs or database queries) to fetch real-time data \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m only after identifying a specific knowledge gap [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **RAG (Retrieval-Augmented Generation)**: The standard for large knowledge bases. Instead of \"stuffing\" a \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m document, the system retrieves only the top-k most relevant chunks from a vector store to populate the \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m context [2][3]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Memory Management\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Agents require sophisticated memory systems to maintain continuity without overflowing context limits: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Short-term Memory**: A sliding window of the most recent interaction turns (e.g., last 10 messages). \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Long-term Memory**: Key facts (e.g., user preferences) are persisted in external databases (Postgres, \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Vector Stores) and retrieved only when relevant to the current query [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Pruning and Trimming**: Algorithms that selectively remove older, less relevant tokens while preserving \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m the semantic core of the conversation history. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Summarization and Compression\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Tool Output Summarization**: Raw outputs from tools (like HTML from a web scraper) are often too \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m verbose. An intermediate step or \"summarizer agent\" distills this data into key points before passing it to \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m the main reasoning agent [2]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Context Compression**: Removing stop words or irrelevant sections from retrieved documents to maximize \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m information density per token. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Prompt Structuring\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Using strict formats within the context window helps the model parse information correctly: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Separation of Concerns**: Clearly formatting \"System Instructions,\" \"User Input,\" and \"Tool Outputs\" \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m (often using XML tags or JSON schemas) prevents the model from confusing data with instructions. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 4. Architectural Frameworks\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Two primary frameworks dominate the implementation of these strategies: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **LangChain** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Focus**: Flexible orchestration and agent workflows. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Approach**: Uses \"Chains\" to pass context sequentially between steps and \"Memory\" classes (e.g., \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m `ConversationSummaryMemory`) to automatically manage history compression and persistence [4]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Strategy**: Emphasizes dynamic tool selection where the agent decides what information to add to its \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m own context. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m **LlamaIndex** \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Focus**: Data indexing and advanced retrieval. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Approach**: specialized in structuring data for LLMs using indices (`VectorStoreIndex`, `TreeIndex`). \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Strategy**: Uses \"Response Synthesizers\" to optimize how retrieved context is presented to the LLM \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m (e.g., `compact` mode to stuff chunks or `tree_summarize` for hierarchical reduction) [4]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 5. Strategic Patterns\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## RAG vs. Long-Context Windows (LCW)\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **RAG**: Preferred for cost efficiency and low latency. It filters noise before the model sees it, often \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m resulting in higher accuracy for \"Needle in a Haystack\" retrieval tasks [3]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Long-Context Models (e.g., Gemini 1.5)**: Best for \"whole-document\" reasoning where global context is \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m required. However, they are more expensive and prone to the \"Lost in the Middle\" effect. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Context Selection vs. Stuffing\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Context Stuffing**: Naively filling the window with all available data. This is generally an \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m anti-pattern as it increases costs and distracts the model. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m * **Context Selection**: The proactive curation of inputs. Techniques include \"Node Postprocessors\" \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m (filtering results by similarity score) and \"Query Transformations\" (rewriting vague user queries to fetch \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m better context) [4]. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;35m## 6. Conclusion\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m Context engineering is the bridge between raw data and intelligent agent behavior. By moving from naive \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m context stuffing to sophisticated dynamic injection, memory management, and structured retrieval, developers \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m can build agents that are both powerful and cost-effective. The choice between RAG architectures and \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m long-context models remains a key trade-off, with hybrid approaches often providing the best balance of \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m accuracy and performance. \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[1;36m#\u001b[0m\u001b[1;36m## Sources\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [1] Context Engineering Guide: https://www.promptingguide.ai/agents/context-engineering \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [2] Context Engineering Strategies (n8n): \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m https://www.theaiautomators.com/context-engineering-strategies-to-build-better-ai-agents/ \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [3] RAG vs Long Context Window Study: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m https://www.legionintel.com/blog/rag-systems-vs-lcw-performance-and-cost-trade-offs \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m [4] LangChain & LlamaIndex Comparison: \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m https://medium.com/badal-io/exploring-langchain-and-llamaindex-to-achieve-standardization-and-interoperabilit \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m y-in-large-2b5f3fabc366 \u001b[34m│\u001b[0m\n", "\u001b[34m│\u001b[0m \u001b[34m│\u001b[0m\n", "\u001b[34m╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "from deepagents.backends.utils import file_data_to_string\n", "\n", "# Convert a specific file to string\n", "file_content = file_data_to_string(result[\"files\"]['/final_report.md'])\n", "show_prompt(file_content) " ] }, { "cell_type": "markdown", "id": "fdcc6784", "metadata": {}, "source": [ "Trace: \n", "\n", "https://smith.langchain.com/public/72d23852-4616-4bcc-8d8a-b0d1905c945b/r" ] }, { "cell_type": "code", "execution_count": null, "id": "2d73925c", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.13.1" } }, "nbformat": 4, "nbformat_minor": 5 }