Self-Evolving Agents with Anytime-Valid Certificates

2026-07-01

AI Summary

This work addresses the challenge that self-evolving agents, by generating their own data, evaluators, and hypothesis spaces, undermine the foundational assumptions of traditional learning theory regarding safety and performance guarantees. To resolve this, the authors propose the SEA (Safe Evolutionary Agent) architecture, which freezes the base model and confines self-evolution strictly to lightweight steering adapters and versioned wrappers. SEA incorporates an always-on validation gating mechanism that, under a fixed error budget, issues auditable certificates to ensure every modification remains safe and controllable. The approach innovatively integrates five closed-loop validators—Best-of-N, micro-step search, self-authored reproducibility oracles, search-layer control, and self-repair—together with a dense-signal mechanism that operates without external scorers, enabling verifiable, non-degrading autonomous evolution. Evaluated on 52 SWE-bench Verified instances, SEA significantly enhances strong base models (e.g., GLM from 24 to 28, GPT from 29 to 34), achieving state-of-the-art performance in 65% of cases, with event logs confirming effective prevention of performance degradation.

Abstract

Self-evolving agents violate the assumption behind most learning-theoretic guarantees: the data, evaluator, components, and hypothesis space are produced by the policy being updated. We present \textbf{SEA}, an architecture that confines self-modification to a small steering adapter and a versioned harness around a \emph{frozen} base model and admits each modification only through an anytime-valid gate that emits an auditable certificate against a fixed error budget. Five loop controllers compose published guarantees; because such gates can only \emph{select} among behaviors the frozen base already produces, five verifier-in-the-loop mechanisms -- best-of-$N$, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair -- supply the dense, grader-free signal the gates require, computed from the issue text alone. On a $52$-instance SWE-bench Verified subset across four base models, base capability is the dominant, confound-free effect, and on two strong base models a deliberate no-op-composite control isolates the suite's contribution at $+4$ and $+5$ (\textsc{Glm}~5.2 $24\to28$; \textsc{Gpt} $29\to34$, the $65\%$ best), with event logs confirming that its mechanisms fire and prevent regressions. Results are single-run on expensive evaluations; confirming run-to-run variance and adapting the per-task algorithm mix are future work.

Problem

• self-evolving agents
• learning-theoretic guarantees
• anytime-valid certificates
• self-modification
• verifiable safety

Innovation

• self-evolving agents
• anytime-valid certificates
• frozen base model
• verifier-in-the-loop
• error budget