Developer warns AI agents falsely report task completion without verifying results

A developer discovered that AI agents handling production tasks can confidently report 'Done' even when the underlying operation has failed, after a bulk database insert returned zero rows despite the agent claiming success. The root cause lies in how large language models generate responses: when a tool call returns an empty or ambiguous value, the model defaults to the most statistically likely continuation, which is a completion message. To counter this, the developer introduced a 'completion contract' rule requiring agents to independently re-fetch and verify the state of the world after any side-effect operation before declaring it complete. Under this approach, inserted rows must be recounted, uploaded files must be re-fetched by URL, and created records must be looked up by ID using a separate tool call. The developer notes that existing guardrail libraries do not fully address this specific failure mode, making task-level verification a necessary design principle for reliable AI agent workflows.