---
type: source-digest
topic: agent-wiki
status: active
created: 2026-04-06
updated: 2026-04-06
source_notes:
  - "[[clippings/26-02-26  ampcode.com  How to Build an Agent]]"
  - "[[ideas/26-03-01 Building Agents]]"
derived_from:
  - workspace-initialize
confidence: high
source_url: https://ampcode.com/notes/how-to-build-an-agent
---

# amp-how-to-build-an-agent

## Why This Matters

This source grounds the workspace in implementation reality. It argues that useful agents are often just an LLM, a loop, tool definitions, and careful engineering discipline, which pushes against overly mystical or framework-heavy thinking.

## Key Takeaways

- A practical code-editing agent can be small and understandable.
- Tool use, state management, and a clear loop account for most of the useful behavior.
- The most impressive-looking agent behavior can emerge from simple parts plus persistence.
- Strong agent systems come from practical engineering and error handling, not secret architecture.
- This is a good counterweight to any temptation to overengineer the knowledge compiler itself.

## Entities

- Thorsten Ball
- Amp
- tool loop
- code-editing agent
- state management

## Related Concepts

- [[research/agent-wiki/concepts/simplicity-before-autonomy|Simplicity Before Autonomy]]
- [[research/agent-wiki/concepts/obsidian-as-agent-ide|Obsidian as Agent IDE]]

## Open Questions

- Which parts of the knowledge compiler could eventually become lightweight scripts instead of prompt-only behavior?
- What minimum toolset would make workspace maintenance substantially better without overcomplicating it?
- How much ceremony is acceptable before the workflow stops feeling lightweight?