Self-Improving Agents Can't Grade Their Own Homework: SEAL Shows One Sealed External Audit Is Enough to Stop Silent Regressions
'Self-Authored Verification Is Unreliable in Heuristic Self-Improving Agents' (arXiv 2607.24300, July 27) formalizes the verifier–deployment gap: when an agent writes both the policy and the test that accepts it, self-assigned scores stay near perfect while real deployment performance degrades. The authors introduce SEAL (Sealed Exogenous Acceptance Loop), which keeps self-authored tests but adds a fixed harness-side audit the agent cannot write or inspect — the agent gets only accept/reject, and the full incumbent state is restored after a clear regression. Failures stratify by capability: weaker agents destroy previously acquired strategies behind easy self-tests, while stronger agents stay stable but still mismeasure the deployment distribution. SEAL beat unprotected baselines across six models and three seeds, and the takeaway is blunt for anyone running a self-improving harness — you need at least one acceptance signal outside the agent's control.
↳ Follow the thread