SEA Self-Evolving Agents with Anytime-Valid Certificates

SEA confines self-modification to a small steering adapter and a versioned harness around a frozen base model, and it admits each modification only through an "anytime-valid gate that emits an auditable certificate against a fixed error budget." The architecture appears in an arXiv paper by Biswa Sengupta submitted 1 Jul 2026 (arXiv:2607.00871).

How does SEA work?

SEA restricts all self-modification to a steering adapter and a versioned harness around a frozen base model, and it enforces changes through anytime-valid gates that issue auditable certificates against a fixed error budget. The paper describes five loop controllers that compose published guarantees, and five verifier-in-the-loop mechanisms that supply the grader-free signal the gates require.

Those verifier mechanisms are listed as best-of-N, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair. Because the gates can only choose among behaviors the frozen base already produces, the verifiers compute a dense signal from the issue text alone, letting the gate select without expanding the base model's hypothesis space.

How well did SEA perform on benchmarks?

On a 52-instance SWE-bench Verified subset across four base models, SEA produced measurable gains on two strong base models: on Glm 5.2 the score moved from 24 to 28 (+4), and on Gpt it moved from 29 to 34 (+5). The paper notes the Gpt result represents the 65% best.

The authors emphasize these are single-run results on expensive evaluations, and they report that event logs confirm the mechanisms fired and prevented regressions. The paper identifies base capability as the dominant, confound-free effect on the suite, and isolates a deliberate no-op-composite control that attributes the suite's contribution at plus four and plus five on the two models cited.

What are the core components and safeguards?

SEA's core design is a frozen base model plus a thin, steerable adapter and a versioned harness. The architecture admits modifications only through gates that produce an auditable certificate, and five loop controllers provide published guarantees around those gates. The verifier-in-the-loop mechanisms compute signals from the issue text alone so the gates can select among existing base behaviours rather than invent new ones.

The paper frames these safeguards as essential for learning-theoretic guarantees because self-evolving agents otherwise break assumptions: data, evaluator, components, and hypothesis space are produced by the policy being updated. SEA confines that feedback loop to auditable, certificate-mediated choices.

Why it matters

SEA addresses a foundational problem for self-modifying agents: how to let a system change its own behavior while preserving external guarantees. By restricting modifications to a steering adapter and requiring anytime-valid certificates tied to a fixed error budget, SEA creates an auditable gate between change proposals and deployment. That design shifts the trust question from continuous model rewrites to verifiable selection among pre-existing behaviors, which matters for use cases that demand auditability and bounded error even when the agent participates in data and evaluator generation.

What to watch

The paper flags two immediate follow-ups: confirming run-to-run variance and adapting the per-task algorithm mix. Future work will need multiple runs to quantify variance and to tune the mix of verifier mechanisms per task to see whether single-run gains hold up and generalize across more base models.

Paper and provenance

Author: Biswa Sengupta. Submission: arXiv:2607.00871, submitted 1 Jul 2026. The abstract and results described here come from that preprint; the authors note evaluations were expensive and single-run, and that adapting algorithm mix and quantifying variance are future work.