Trust execution proof, not an AI agent’s claim that a task is done
In real services, an can report that it sent an email, changed a customer record, or created a support ticket even when the required tool never ran. The run may show no error and appear successful, so the failure becomes visible only when the expected result does not happen. Asking the model whether it really performed the action is not a reliable check because it may confidently confirm a step it skipped.
Repeating the prompt does not solve this problem. should require an showing that the occurred during that step and returned evidence of running. If the record does not match the claimed action, the result is unknown rather than complete; an empty or null result should not count as success either.
The should move forward only after receiving valid execution proof.
Key points
- Do not accept an agent’s own claim as proof.
- Check that the and its result appear in the execution record.
- Treat a claim without evidence as unknown, not complete.
- Do not count an empty or null tool result as success.
- Advance the only after valid execution proof is available.