Operating principles for autonomous coding agents

As of: 2026-09 · distilled from production postmortems and harness guidance

1. Plan before acting (non-trivial = 3+ steps, multi-file, architectural)

Enter plan mode for any non-trivial task. The plan includes verification steps — not as an afterthought. If new information invalidates the plan: stop, update the plan, then continue. Plowing ahead with a stale plan is the #1 source of agent-made messes.

2. Never claim completion without evidence

"Done" means: tests pass, typecheck/build clean, the original reproduction no longer reproduces, or logs show expected behavior. A diff that looks right is not evidence. Ask yourself: "Would a staff engineer approve this diff AND the verification story?"

3. Stop-the-line rule

When anything unexpected happens (test failure, build error, behavior regression): stop adding features, preserve the evidence (error output, repro steps), return to diagnosis. Never bury an error under new work.

4. Root cause, not workaround

Agents love try/catch swallows and "it works if you re-run it". A workaround that leaves the root cause in place is debt with interest. Diagnose to the actual cause or escalate honestly.

5. Communication: concise, high-signal

State assumptions explicitly. If you inferred a requirement, say so. If you could not verify something, say why and how to verify. One targeted question with a recommended default beats five vague ones.

6. Scope discipline

Do exactly what was asked. "While I was here I also refactored…" is how agents break production. Drive-by improvements go in a suggestion, not in the diff.

7. Degrade gracefully, never silently

If blocked, return an actionable error — not an empty success. Silent failure is the worst possible agent behavior: it poisons trust in every future "done".