The Autonomous Agent Operating Manual

Built and open-sourced by the team behind Millennium AI

A battle-tested operating manual for running a coding agent autonomously on real production systems — distilled into one file you can drop in your repo.

We run Claude Code autonomously against a live crypto-trading platform: it edits code, runs tests, commits, and deploys with minimal hand-holding. The single thing that made that safe instead of terrifying was a tight, opinionated CLAUDE.md. This repo is the generalized, secret-free version of it.

👉 The whole thing is one file: AGENTS.md. Copy it into your repo as CLAUDE.md, AGENTS.md, or .cursorrules and trim to taste.

---

Why you want this

Coding agents fail in predictable ways: they edit code they don't understand, smuggle mock data into production, blow time/token budgets, break shared contracts they didn't trace, and confidently claim things are "fixed" without checking. This manual is a fence built around exactly those failure modes.

It's organized as Four Prime Directives (the 20% you must keep) plus practical guardrails:

1. Map First — understand the system end-to-end before editing.
2. No Fake Data in Production Paths — mocks live in tests, never in prod.
3. Respect Budgets — time, tokens, money, latency all have limits.
4. Trace Downstream — follow every shared-contract change to the user.

…followed by sections on autonomy vs. irreversible-action guardrails, git discipline, errors as learning, token/cost optimization, a file-based memory pattern, when to stop and ask a human, and a pre-commit checklist.

How to use it

# in your project root
curl -O https://raw.githubusercontent.com/MrMaca11an/autonomous-agent-md/main/AGENTS.md
mv AGENTS.md CLAUDE.md     # or keep as AGENTS.md / .cursorrules

Then delete what doesn't apply to you and keep the principles. The structure matters more than the exact wording.

What makes it different

• Written for real autonomy, not demos. It assumes the agent will act without asking — so the guardrails are about what it must never do.
• Incident-first. Stabilize the service before refactoring.
• Memory built in. A small file-based "LLM wiki" so the agent stops relearning your project every session.
• Cost-aware. Token/compute optimization is a first-class section, not an afterthought.

Contributing

Found a failure mode the manual doesn't cover? Open a PR with the principle and a one-line rationale. Keep additions short — this document's value is in being read, which means staying short.

License

MIT. Maintained by the team behind Millennium AI.