# Agent Wiki

## Purpose

Build a research workspace for the idea that LLMs can maintain a durable markdown knowledge base, and use that pattern to sharpen how agent systems should be designed, scoped, and integrated into this vault.

## In Scope

- Notes and clippings about agent architecture, agent workflows, and LLM-maintained knowledge systems
- Material about simple vs. complex agent design
- Material about Obsidian or markdown-first environments as the operating surface for agent work
- Questions about how this vault should instantiate the Karpathy-style wiki pattern

## Out of Scope

- General AI market commentary unless it directly changes the design of agent systems
- Broad product strategy for all AI tools in the vault
- Full-vault compilation or a global ontology for every note type
- Implementation details for a specific code agent unless they illuminate the knowledge-system model

## Current Thesis / Working Questions

- A durable markdown layer between raw sources and downstream writing is more valuable than repeatedly querying raw notes.
- Good agent systems should default to simple, inspectable workflows before escalating to autonomy.
- Obsidian can function as the human-facing IDE for this system if the LLM owns local compilation and the human owns scoping, judgment, and promotion.
- The workspace should behave as a governed interpretation layer, not as a silent authority store.
- What rules make a local-first, shared-when-earned knowledge system trustworthy over time?

## Output Goals

- Durable synthesis notes that explain the agent-wiki pattern
- Comparison outputs contrasting compiled knowledge systems with ordinary RAG or chat memory
- Inputs to future `ideas/` notes and posts on agent workflows, software interfaces, and knowledge compounding

## Promotion Candidates

- Compiled knowledge layer
- Simplicity before autonomy
- Obsidian as agent IDE
- Governed interpretation layer

## Known Tensions

- A persistent wiki compounds value, but it also introduces maintenance and ontology drift risk.
- Shared concepts reduce duplication, but premature promotion can create brittle abstractions.
- Simplicity is desirable, but some useful workflows may still require stronger tooling over time.
- A compiled layer increases continuity, but it also creates delayed attack surfaces if poisoning, stale synthesis, or provenance loss are not contained.