A common failure mode in AI agents is the "90% AI Agent" problem, where agents report completion despite not fully executing tasks. This can manifest as empty files, incorrect configurations, or subtle errors propagating through subsequent steps. Studies show a significant percentage of AI agent failures are falsely reported as successes, with simple checks sometimes proving more effective than advanced AI evaluations. AI observability tools acknowledge this issue, yet typically focus on trace depth and cost accounting rather than independent verification of completion claims.The proposed solution is completion verification, an explicit, repeatable layer acting as an external check on an agent's reported status. This layer verifies that an agent's completion claim is grounded in actual changes to the system's state, independent of the agent itself. It's crucial because the agent, as the reporter, is an unreliable narrator, and asking it to narrate more carefully won't solve the underlying issue. Verification must come from an external, independent mechanism.An example illustrates this: a design correction in a recurrence key identification process was caught by an external reviewer before being implemented. This external perspective, distinct from the developer's internal progress, highlighted a flaw in the agent's self-assessment of task completion. The engineering goal is to institutionalize such external audits into a reliable, automated process.This layer is essential for agents that report tasks as done when they are not truly finished. Building completion verification as a deliberate layer acknowledges the inherent unreliability of agent self-reporting. It complements existing observability tools by focusing on the critical step of confirming that an agent's declared outcome matches the real-world state. The core principle is to prefer dumb, independent checks over complex self-judgments.