Agent-Checks-Agent: Using Secondary Agents to Verify Task Completion Claims

Solution

Deploy a dedicated verification agent that runs after every primary agent session. The verification agent receives: the original task specification, the list of files modified, and the claimed completion status. It independently re-examines the codebase changes against the task spec without access to the primary agent reasoning or tool outputs. It checks acceptance criteria independently, looks for obvious gaps in the implementation, and reports a separate confidence score. The two agents use the same model but different system prompts and zero shared context. Disagreements between primary and verification agent claims trigger a human review step rather than auto-approval. In OpenClaw, implement this as a post-task hook that always fires on session completion. Track agreement rate over time as a signal of primary agent reliability.