How agents.md fits Compound Engineering Compound Engineering is built on file-based knowledge accumulation: instead of keeping everything in chat history, project context lives in files the workflow reads and writes [^38395538]. The agents.md file sits in that layer as the agent-facing instruction file — the place where the guidance and context agents need to do their work is captured, so that multi-agent, multi-model sessions start from a shared source of truth rather than from scattered conversation. Why it matters in a multi-model world The v3.20 release discussion frames the whole system around a multi-model workflow, where different models handle planning, work, and review, and the goal is to accumulate knowledge rather than burn tokens [^38395538]. A file like agents.md is what makes that possible: because instructions are persisted in the repo, any model or skill in the loop — from core planning skills to newer ones like ce-pov and ce-explain — can consult the same accumulated project knowledge [^38395538]. What you will take away - agents.md belongs to Compound Engineering's file-based knowledge layer, holding agent-facing instructions and context.- Persisting that guidance in a file lets multi-agent, multi-model sessions share one source of truth.- Keeping agent guidance in the repo supports the compounding plan-work-review cycle instead of relying on chat history [^38395538].