The gap between 'an agent that can write code' and 'an AI employee that can actually get work done' is often not model capability but whether it starts with your team's context on board. Codex's answer is AGENTS.md: a project-level instruction file placed in the repo root. The official docs split 'Agent configuration' into layers — AGENTS.md carries project-level instructions and context, Subagents package specialized sub-tasks into reusable sub-agents, Speed controls response pacing, and Rules constrain behavior. The essence is 'externalized conventions': pulling 'how this repo builds, tests, and names things' out of every chat's prompt into a file, so any session and any collaborator can reuse it stably. OpenTelemetry's spec repo maintains its own AGENTS.md (and CLAUDE.md), showing this is already a default open-source convention.
For Zero to Codex learners, AGENTS.md is the first 'engineering' move to master: after installing Codex, the first thing isn't letting it write code freely — it's writing build commands, test commands, and directory structure into AGENTS.md. The course contrasts it with ad-hoc prompt instructions so learners feel the reproducibility that 'externalized conventions' bring. For Zero to Pro Agent Harness learners, AGENTS.md is a humble prototype of 'org-wide configuration delivery' — FIIT.AI's AI employees receive org-wide conventions from the cloud and new employees inherit them automatically, which is essentially upgrading AGENTS.md's idea from 'one repo' to 'org level'. Understanding this layer reveals why 'context engineering' is closer to production than 'prompt engineering'.