Context engineering in 2 moves: one short AGENTS.md root file plus docs loaded on demand
Al_Grigor · x · 2026-10-27
Al Grigor shares his two-rule context engineering discipline for coding agents, solving the cold-start problem of every new session:
- One root file: all instructions live in a short AGENTS.md; CLAUDE.md is a single line pointing to it, so Claude Code and other agents read the same file.
- Keep the root file short, use it as a map: it links to focused docs — process docs (docs/process.md with PM/Engineer/QA role files), testing guidelines, UI/design rules, API conventions. The agent loads only what the task needs: tests → testing docs, UI → design rules.
Start small: begin with the process doc and add role files as needed.
More from coding & agent
- Box explains its Deep Agent architecture: parent-child agents over a LangGraph orchestrator — LangChain · 2026-10-28
- How Box built enterprise content agents on Deep Agents with a parent/child architecture — LangChain · 2026-10-28
- How do you inject instructions mid-run? Dev struggles with stage summaries agents don't trust — tinker_bob · 2026-10-28
- Harrison Chase to demo agents that catch their own mistakes and write the eval and fix — LangChain · 2026-10-28
- Parallel launches Parallel Decisions API at $0.001 per request with 5,000 free calls — schwentker · 2026-10-28
- Together breaks down the MIGHT stack: composable open-source AI from models to MCP tools — togethercompute · 2026-10-28