Self-Evolving Agents Gain Anytime-Valid Certificates for Reliability.

Self-evolving agents, which modify their own policies and components, challenge traditional learning-theoretic guarantees because the very system being updated also produces its own data and evaluation. This research presents SEA (Self-Evolving Agents), an architecture designed to manage this complexity by confining self-modification to a small steering adapter and a versioned harness around a frozen base model. A core innovation of SEA is the "anytime-valid gate," which only permits modifications if they come with an auditable certificate, ensuring adherence to a predefined error budget. To provide the necessary dense, grader-free signals for these gates, SEA incorporates five verifier-in-the-loop mechanisms. These mechanisms, including best-of-N selection, micro-step search, and self-authored reproduction oracles, compute signals directly from the issue text, enabling robust self-repair and preventing regressions. Evaluations on a SWE-bench Verified subset across multiple base models show that while base capability is the primary factor, SEA's mechanisms contribute a measurable improvement, preventing performance degradation. The results, though single-run on expensive evaluations, confirm that the system's mechanisms actively fire and prevent regressions, pointing towards a more reliable path for self-modifying AI.