Agent Self-Evolution · Statistical Guarantees

The loop that
checks itself.

First surfaced in the Tandemly Briefing for 2026-07-12.

Most self-improving systems have no built-in test for whether they actually improved. They update, and they hope. SEA changes that: a frozen base model accumulates small adaptations from verifier feedback and then runs a statistical gate before any change takes effect. The gate can answer at any point during evaluation. If the numbers do not support the modification, the modification does not land.

Core concept
Anytime-valid admission control: a statistical gate that can reach a valid conclusion at any point in an evaluation, controlled against a pre-specified false-positive budget, before any self-modification takes effect.
scroll to explore

Self-improvement that
can't verify itself is just drift.

Standard approaches to improving a deployed agent are either expensive or ungated. Neither option is appropriate for a production system where every failed task has a real cost.

Retraining or fine-tuning a model to improve one capability is expensive and slow. Self-play and reinforcement approaches are faster, but they introduce a reliability problem: how do you know a model that completed a few extra rollouts actually got better, rather than just got lucky? Sequential A/B tests can answer this, but only after a predetermined number of trials. If the signal is clear earlier, you have wasted evaluation time. If the signal is not yet clear at your predetermined horizon, you have committed to an answer anyway.

Neither outcome is good in a production agentic system. Early stopping at significance would let you ship improvements faster or reject failures sooner. Committing at a fixed horizon means you accept the false-positive rate at that one point in time rather than across the full evaluation trajectory.

There is also a deeper problem with existing self-improvement approaches: they do not separate the improvement target from the base model. Every update changes the same parameters that handle tool calls, state tracking, and structured output. Small improvements in one capability can silently degrade another. And if something goes wrong, you cannot revert the capability without reverting everything else.

The question this paper asks

Can a self-evolving agent accumulate improvements from its own verifier signals, gate each modification behind a statistically valid test that controls the false-positive rate at any stopping point, and do all of this without touching the base model that everything else depends on?

Freeze the core.
Adapt at the edges.

SEA separates the improvement target from the production model. The base model is frozen. A small learned steering adapter receives all modification. The adapter accumulates signal from five verifier loops that evaluate the agent's outputs during normal task operation.

The frozen-base architecture is the first design choice. The agent's base weights never change. A lightweight steering adapter is trained on top, accepting modification while the base model stays stable. Reverting an adaptation is inexpensive. Re-training a base model is not. This also means the base model's general capabilities are protected from the gradient noise of narrow-task improvement.

The adapter accumulates signal from five verifier loops that operate during task execution. Each loop targets a different structural weakness in agentic reasoning.

1
Best-of-N sampling
Sample N completions for the same sub-problem; the one that passes verification becomes the training signal for the adapter. Isolates what a correct attempt looks like from what a typical attempt looks like.
2
Micro-step search
Search over sub-problem decompositions before committing to an approach. The decomposition that leads to a verified completion wins and produces training signal. Addresses the commitment problem in long-horizon tasks.
3
Reproduction oracles
Check whether completed code passes a test suite generated independently of the solution. Failure produces a correction signal. The oracle operates at the sub-step level, not just on final task outcomes.
4
Search-layer control
Manage the depth and branching of search steps explicitly, so the agent does not get stuck in loops or commit too early to a path that precludes better options later.
5
Self-repair
When a step fails, route the failure description back to the agent with a correction prompt. The repaired attempt becomes additional training signal. Produces learning from what actually went wrong, not just from what a correct answer looks like.

Before any modification lands, the system runs an anytime-valid statistical test against the candidate change. Unlike a fixed-horizon t-test, anytime-valid tests can produce a valid conclusion at any point in an evaluation: if significance is reached at step 20 of a 100-step evaluation, the test can stop there. If it is not reached by step 100, that is also a valid result. The statistical guarantee covers the false-positive rate across all possible stopping times, not just at one predetermined endpoint.

Modifications that clear the gate receive an auditable certificate: a record of the evaluation trajectory, the test statistic, the stopping point, and the conclusion. The certificate is the evidence trail for any downstream review.

What "anytime-valid" means

Anytime-valid tests use e-values or sequential likelihood ratio tests rather than classical p-values. The false-positive rate is controlled regardless of when you look at the data, not just at a pre-specified endpoint. This means you can monitor continuously and stop as soon as you have enough evidence, without inflating Type I error. Classical testing controls error only at one point; anytime-valid testing controls it everywhere.

Both baselines grew.
GPT grew more.

Evaluated on a 52-instance subset of SWE-bench Verified, a benchmark for coding agents that requires resolving real GitHub issues. Both backbone models improved under SEA, with GPT showing the larger absolute gain.

GLM 5.2 + SEA
28
+4 from 24 baseline
GPT + SEA
34
+5 from 29 baseline
Evaluation scope
52
SWE-bench Verified instances

The baseline conditions represent the same models without the steering adapter and without the verifier loops. Both gains are positive and direction-consistent across backbones, which suggests the benefit comes from the SEA structure rather than from properties of a specific model. Each additional resolved instance on SWE-bench Verified represents a genuinely novel code change that passes an independently written test suite, so the gains are not artifacts of evaluation design.

The gains are also sub-additive: verifier signals do not compound without bound. The steering adapter has a capacity limit, and the gate exists precisely to prevent modifications that look like improvements but have not cleared a sufficient evidentiary bar.

Scope and limitations

This is a single-run evaluation on a 52-instance subset. Run-to-run variance has not been characterized. The steering adapter architecture is described but has not been independently replicated. The anytime-valid gate is the paper's central statistical claim; the formalism is sound, but the practical benefit of early stopping over a well-designed fixed-horizon test in this specific setting has not been quantified separately. These limitations do not undercut the direction of the results, but they do mean the absolute numbers should be treated as directional rather than definitive until independently confirmed.

What changes,
and what stays fixed.

SEA contributes three durable patterns: separate the improvement target from the base model, gate every modification with a statistically valid test, and verify at the sub-step level rather than only at task completion.

1
Separate the improvement target from the base model
When you need an agent to get better at a specific capability, accumulate that improvement in an adapter rather than re-training the weights that everything else depends on. The base model stays stable; the adapter absorbs the change. Reverting an adaptation is inexpensive; re-training a base model is not. This separation also protects unrelated capabilities from gradient noise.
2
Use anytime-valid statistics for continuous model evaluation
Classical hypothesis tests require committing to a sample size before you start. Anytime-valid tests let you check at any time and stop when you have enough evidence, without inflating the false-positive rate. This matters most when evaluation is expensive or when early stopping would let you ship an improvement or reject a failure sooner.
3
Gate self-modification with auditable certificates
Any system that modifies its own behavior needs a paper trail. Generating a certificate at each modification point (the evaluation trajectory, the test statistic, the stopping point, the conclusion) creates a reviewable history. When something goes wrong downstream, you can trace which modification introduced the change and whether it passed its gate at the time.
4
Verify at the sub-step level, not only at task completion
All five SEA verifier loops operate on sub-problems, not final task outcomes. Reproduction oracles and search-layer control provide signal mid-task. Task-level evaluation can mask sub-problem failures if the agent compensates later. Sub-step verification catches these failures earlier, when the steering signal is still useful.

How to apply
these patterns.

You do not need the full SEA system to benefit from its core ideas. Each of the three main patterns can be piloted independently.

1
Apply the frozen-base pattern now
If you are fine-tuning a model for a specific downstream task, consider training a small LoRA or adapter on top of a frozen checkpoint rather than updating all weights. The base model retains its general capabilities; the adapter specializes. This is available today with standard parameter-efficient fine-tuning libraries (PEFT, Hugging Face). You can evaluate the adapter in isolation and roll it back without touching the base.
2
Pilot anytime-valid tests in your evaluation pipeline
Replace one A/B test or model comparison with an anytime-valid equivalent and compare how often you would have reached a valid conclusion earlier. Python implementations exist in the sequential testing literature (see Ramdas et al. 2023, "Testing by Betting"). The cost of piloting is a single pipeline substitution; the upside is evaluation runs that can stop as soon as they have an answer.
3
Build a certificate template before building the gate
Define what a "passed modification" certificate should contain before you instrument the gate: the evaluation run ID, the metric being tested, the stopping criterion, the result, and a timestamp. Having the schema first keeps the audit trail useful rather than a post-hoc justification that grows to fit whatever the system already records.
4
Read the paper
Sengupta, B. (2026). Self-Evolving Agents with Anytime-Valid Certificates. arXiv:2607.00871. The formalism for the anytime-valid gate and the architecture of the steering adapter are described in full.