Self-Evolving Agents with Anytime-Valid Certificates
2026-07-01 • Artificial Intelligence
Artificial IntelligenceComputation and Language
AI summaryⓘ
The authors identify a problem with self-evolving agents, where the data and evaluation change as the agent updates itself, breaking typical learning assumptions. They propose SEA, a system that limits self-changes to a small part while keeping a main model fixed and only allows updates if they pass strict checks based on a fixed error limit. These checks rely on multiple verifier methods that test and select behaviors already present in the base model, using only the problem description. Their tests show improved performance when these mechanisms are added, and event logs confirm the system prevents negative changes. Future work will address repeating tests and adjusting methods per task.
self-evolving agentslearning guaranteesSEA architecturefrozen base modelanytime-valid gateauditable certificateverifier-in-the-loopmicro-step searchself-repairSWE-bench
Authors
Biswa Sengupta
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.