Skip to content

AI agents

AI agents may write issues, code, tests and docs for ritueel. Their work is reviewed like any other, and they follow the same pages as people. AGENTS.md at the repository root is the short form of these pages for an agent, with the page to read before each kind of change; CLAUDE.md loads it for Claude Code.

  • A person owns the contribution. That person has read it, understands it and answers for it.
  • Mention the assistance in the pull request.
  • An agent does not decide scope and does not touch governance files unless asked. It presents options with pros, cons and a recommendation, and the maintainer decides (Decision records).
  • Output that is plausible but unverified is not a contribution. Run it, test it, and say what you ran (Tools, checks and verification).
  • Low-effort bulk pull requests may be closed without review.

A commit or a pull request that an agent wrote keeps the attribution lines the tool adds, such as a Co-Authored-By trailer. A DCO sign-off, where one is asked for, is the human contributor’s legal statement: an agent never adds one on its own.

An agent does not work around a safety check, a permission prompt or a refused command. If one blocks the work, the agent stops and reports what blocked it and what it was trying to do.

When several agents work on one change, each has one role, and the roles stay separate. There is no single combined “reviewer”.

Role May change Does
implementer source code, including the site’s code in site/ builds the change
security nothing: read-only reviews for faults and reports findings; never edits
qa tests and test data writes and runs the tests that prove the change
docs docs, README, CONTRIBUTING.md, AGENTS.md, CLAUDE.md, the agent files in .claude/ keeps the pages and the agent wiring in step with the change
devops CI workflows in .github/workflows/, the repository checks in tools/checks/, the Justfile, devbox.json, release configuration keeps the checks and the release path working

The review agent of GOVERNANCE.md is a different agent from the one that wrote the change. For Claude Code the roles are defined in .claude/agents/, and the common tasks have skills in .claude/skills/ that point back to these pages.

An agent ends its work with a report:

  1. Verdict. What was done or found, first.
  2. Numbers. Measurements, counts, check results, with the commands that produced them.
  3. Changes. Every change it made to any machine, account, branch or device, and whether it was undone.
  4. Open questions. Each as options with pros and cons, and a recommendation.

A reviewer, such as the security role, reports each finding with the file and line, what could go wrong, how likely it is, and a test that would show it.