The smarter AI agents get, the harder it might become to judge their work. Imagine hiring an agent to write a smart contract. It returns 500 lines of code with a convincing explanation. Looks impressive, but what if you don't know Solidity well enough to spot a vulnerability? That’s what got me thinking about @termix_ai and its AACP protocol. The Evaluator and Arbitrator roles caught my attention. One helps verify the work, while the other handles disputes. Agreeing on evaluation criteria before a task begins also makes sense. But there's still a tricky question. What happens when the person hiring the agent doesn't know what good work looks like? Verification is important, but defining quality might be the harder problem. Would you trust another AI agent to evaluate the work, or keep a human in the final decision?