Google Cloud AI Research's answer to overfitting in automated harness evolution. Methods that let an LLM propose and keep edits to an agent's harness (prompts, control flow, tools, memory, context management) amount to recursive self-improvement at the agent-system level, but because every round re-scores candidates on the same finite evolve set, the harness can memorize that set: large in-distribution gains shrink or vanish on other benchmarks. RRSI leaves the whole harness editable and regularizes the search instead. On the proposal side, a cosine-annealed budget caps how many independent edits one candidate may bundle (the paper's L0 analogy), a per-edit history keeps the proposer from redrawing hypotheses that already failed, and when progress stalls within the noise band part of the budget goes to components not yet touched. On the selection side, an LLM critic rejects edits that encode benchmark-specific content (task names, answers, task-specific values) before they are scored; a candidate must stay above the best score so far minus a noise band estimated from repeated runs of the base harness; extra policy tokens must be paid for by a proportional score gain (the Ridge analogy); and components that stop producing measurable gains are pruned (the Lasso analogy).

Claude Opus 4.8 serves as the frozen policy, proposer, analyst and critic. In each of three domains the harness evolves on one suite and runs unchanged on held-out ones: Terminal-Bench 2.1, then SWE-bench Verified; Harvey LAB (120 evolve and 40 held-out legal tasks), then JobBench, GDPval and APEX-Agents; EngDesign, then Frontier-Eng. Against Meta-Harness, AHE, TTHE and HarnessX, all run from the same base harness with the same candidate budget, every baseline gains more on the Harvey LAB evolve split (up to 93.0, against RRSI's 90.5). Out of distribution the ranking flips: RRSI averages 43.6 across JobBench, GDPval and APEX-Agents against 39.7 for the unevolved harness, while the strongest baseline, Meta-Harness, adds 0.9 and AHE and TTHE end below where they started. Unregularized evolution reaches the highest evolve score (92.8) but only 40.3 out of distribution, at 3.80 million policy tokens per trial against RRSI's 2.42 million (1.56 million for the base harness). In coding, RRSI lifts Terminal-Bench 2.1 from 74.2 to 80.2 with Opus 4.8 and from 64.6 to 78.7 with Gemini 3.5 Flash as the policy (the paper's 14.1-point headline), carrying gains of 1.8 and 2.2 points to SWE-bench Verified; the Gemini-evolved harness also lifts Gemini 3.1 Flash Lite, which the search never used, from 11.2 to 14.6. Frontier-Eng gains 4.3 medal points, and no held-out split regresses.

Eleven of the 14 authors are at Google Cloud AI Research, including first author Peng Xia (UNC-Chapel Hill, working as a Google student researcher) and senior authors Tomas Pfister and Chen-Yu Lee; the other co-authors are from UNC, Stanford and Washington University in St. Louis. The same group released EnvHarness a month earlier. The Apache-2.0 code runs each candidate in its own git worktree and logs every edit's hypothesis and measured effect; it had about 1,000 GitHub stars by September 30, 2026. SoL-Pi, NVIDIA's harness auto-research posted four days earlier, avoids the same overfitting by keeping its benchmark out of the search, and the AI2/UW study RRSI cites measured how little harness-evolution gains transfer.

A sibling project followed on October 6. Harness² (arXiv 2610.14224; Chen-Yu Lee, Tomas Pfister, Rujun Han, Yanfei Chen and Han Yu are authors on both papers) takes on the case RRSI does not cover, open-world tasks that have no evolve set and no ground-truth verifier. At the task level a contrastive Propose–Probe–Compose step proposes edits to eight harness components, probes them on the same task, and composes a refined harness from how the executions differ. At a second level, a persistent "improvement harness" collects the lessons learned for later tasks. With GLM-5 and Gemini 3.5 Flash, the authors report that a single step adds up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB's partial pass rate. The Apache-2.0 code, which supports OpenCode and Codex executors, drew 18,700 GitHub stars in its first 18 days, far more than RRSI.

Alibaba Research's Recursa (arXiv 2611.04035, November 2) applies the same idea to what a system knows rather than to its control flow. The authors are Erkang Zhu, Yin Lin, Xiaorong Shi, Xuanrui Lin and Alibaba CTO Jingren Zhou. Like RRSI, it measures on Harvey LAB, and like RRSI it leaves the model frozen, without a stronger teacher or human experts. What it learns is a knowledge repository of domain skills, rules for when to apply them, task memories and reflections. That repository is first bootstrapped from the domain corpus and from practice on self-generated tasks. It is then refined by a recursive learner that splits the training tasks among child learners, validates and merges their revisions, and stores reflections on how they learned. On the Firm Knowledge split (150 training and 100 held-out tasks), the headline test all-pass rate rises from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash. Bootstrapping alone adds little, and the merge step accounts for most of the gain. A single learner given the same budget peaks at 43.0%. Two caveats come from the paper's own tables. The scores use rubrics the authors revised (on the original rubrics the DeepSeek peak is 54.0%). And the test score does not rise steadily across rounds: 55.0 after round two, 43.0 after round three, 48.0 after round five. The headline is therefore the best round on held-out tasks, not the final one. The paper frames the repository as a first step toward a fully self-improving system that would also retrain the model and the harness, and it does neither.

Paper

Authors: Peng Xia · Rujun Han · Zifeng Wang · Yanfei Chen · Yufan Zhuang · Yoonho Lee · Chengsong Huang · Han Yu

Library

Language Python
License Apache 2.0
agentsagent-harnessself-improvementcodingopen-sourceresearch

Related