Agentic Engineering Standards

An agent that writes code is a fast junior with infinite stamina and no memory of consequences. It will produce plausible diffs at 3am and, unsupervised, plausible damage too. The fix is not a better model - it is the same engineering discipline good teams already have, enforced where the agent cannot bypass it: in branches, reviews, CI and merge rules.

• Agents work on feature branches only - direct commits to main are disabled by branch protection, not by convention
• Every agent change arrives as a merge request with a description a human wrote or reviewed - the 'why' is never generated silently
• CI must pass before merge and the merge button requires a human - pipeline green is evidence, not permission
• Diffs are reviewed like any contributor's: secrets, license headers, dependency additions and generated files get explicit attention
• Secrets never enter agent context - credentials live in the secret store; if a value reaches a transcript it is rotated as an incident
• Agents verify against the real system, not assumptions - a claim of 'deployed' or 'merged' is confirmed by reading actual state
• Lockfile and dependency updates are owned by the dependency bot (Renovate) - agents do not compete with it on the same branch
• Destructive operations need explicit human instruction - force-push, deletes and history rewrites are never inferable from context
• Generated artifacts are reproducible - a script in the repo rebuilds them, so the diff can be audited rather than trusted
• Post-merge state is verified - the applied revision in deployment matches the merged commit, not just a green pipeline

Speed without audit debt. Our agents ship several merge requests a day under these rules, and the review load stays human-sized because the rules catch the failure modes early. A team that adopts the checklist keeps the velocity and loses the weekend incident. That is the entire trade - there is no version where the discipline is optional and the outcome is the same.