Don't verify the agent. Verify the state.

Building an AI agent that checks its own actions proved we overlooked a key challenge: how to confirm that an action truly occurred. Checking the agent's report doesn't solve the problem, as it merely extends the trust chain. Every step trusts the output of the previous step, with none verifying the world directly.

Instead, we found success by examining system changes against our expectations. When sending an SMS to Sarah, we checked the sent folder for a message and compared timestamps. If both checks passed, we reported confirmation. This verification is independent of the agent, relying on ground truth outside of it.

However, some actions lack observable state changes. For these, the honest response isn't success or failure, but UNRESOLVED. We implemented this status in StareBrain as a distinct first-class signal, rather than a temporary placeholder. The confirmation screen acknowledges whether the state is observable or not.

This approach is crucial for AI agents. An agent confidently declaring success on an unverifiable action is worse than one that remains silent. Silence signals uncertainty, while false confidence removes that signal entirely. By placing the verification layer outside the trust chain, we achieve this goal for most actions. UNRESOLVED handles the rest honestly.

StareBrain, an Android AI agent, allows users to specify tasks, preview the action, and confirm before execution. Currently, the pre-launch waitlist is open.