Why AI agents report unfinished work as done

AI agents can report a task as completed even when it hasn't been fully finished. This occurs when the tool returns a success status, but downstream systems or operations haven't actually been updated. The agent needs to verify the expected final state, not just the interface success. It requires a contract outlining the expected final state, including verification methods and recovery paths.

The contract should exist before the agent runs the tool, as it cannot validate the intended result by itself. An agent should only claim a task is done after a separate read confirms the workflow reached the desired state. For example, a refund might require a refund record, ledger entry, and visible status, while a deployment needs the correct account, resources, and healthy service checks.

Trusting an agent that claims a system is still on is risky, as it may have disabled monitoring controls. Verifying through an independent source, such as querying the system of record, is crucial to ensure the action worked as intended. The verification should provide structured evidence, including the target ID, observed state, expected state, and verdict.

If the verification finds the expected state, the agent can mark the work as complete. If not, it should initiate a recovery process or alert a human operator. Treating a timeout as an unknown state is essential, as it could result from various scenarios, and labeling it as a failure may cause the agent to retry and potentially duplicate actions.