This loop tackles the most common failure: an agent that says “done” when it isn’t.
Before any work, it writes a completion contract listing each requirement and the concrete evidence that proves it, then executes item by item.
The cycle implements one requirement, attaches its proof (a passing test, command output, a diff), and repeats until every requirement has evidence.
No evidence, not done.
When to use
- When “done” needs to be objective and auditable.
- For multi-requirement tasks that tend to end up half-finished.
- When you want evidence, not the agent’s self-report.
How to use
Run the goal asking it to first write the completion contract, confirm it, implement requirement by requirement, and attach the evidence for each before declaring success.