Agent Self Check

Agent self-check

The cheapest way to raise agent quality is to add a structured self-check before declaring done. Most "the agent shipped a regression" reports trace back to a missing self-check step.

When to run

• Before posting "done" / "complete" in any planning task you own.
• Before closing a Jira / GitHub Issue.
• Before sending a customer-facing answer that involved tool calls.

The self-check

Ask yourself each question, IN ORDER, and answer it explicitly in your reply (or in the task comment). Do not skip.

1. Did I actually run the change? Either a unit test, a smoke run, or a real call to the tool I claim to have used.
2. Did I check for the silent failure mode? What is the worst thing that could be wrong with my output that LOOKS right? Did I check for it?
3. Did I update the related docs/playbooks? If the change is permanent or recurring, write it down. Use sks_docs_propose_patch or initiative_file_upload.
4. Did I leave an observability hook? For code: a log line / metric / alert. For ops: a planning comment with the timestamps and values.
5. Did I follow my own playbooks? Did I trigger production-validation, release-readiness-checklist, etc. when the work calls for them?
6. Did I tell the right people? A handoff packet if someone else continues; a customer-facing summary if the user is waiting.

When to add a step

If you find a gap in the self-check, fix it in the same turn — do not declare done. If the gap requires a tool you don't have, run discover-capability-and-request-access and pause.

Anti-patterns

• Self-checking by re-stating the original goal.
• Skipping the silent-failure-mode question.
• Marking done with TODO comments still in the diff.