Google RRSI stops agents from gaming their own tests

Google Cloud AI Research, with partners at UNC Chapel Hill, Stanford, and Washington University in St. Louis, published RRSI: Regularized Recursive Self-Improvement of Agent Harnesses. The paper tackles a quiet failure mode in agent systems that rewrite their own prompts, tools, and control flow from test feedback.

What went wrong before

When an agent harness keeps optimizing against the same evolve set, scores on those tasks rise while gains on new tasks shrink or vanish. The search can memorize benchmark quirks, chase evaluation noise, and pile on complexity that never helps outside the training suite.

What RRSI changes

RRSI leaves every harness component editable, but regularizes the search itself. A shrinking edit budget limits how many changes land in one round. A critic blocks proposals that hardcode task names or answers. A pruner removes parts that stop earning their keep. Costly edits only stick when they buy a clear score gain.

Results that matter

Across eight benchmarks in coding, agentic office work, and engineering design, with Claude Opus 4.8 frozen as the backbone, RRSI gained up to 14.1 points on the evolve split and up to 4.7 points on five out-of-distribution benchmarks. The regularized harness also used about 30 percent fewer policy tokens than unregularized evolution. No held-out split fell below the unevolved baseline.

A harness evolved with Gemini 3.5 Flash still helped a weaker Gemini 3.1 Flash Lite model, lifting Terminal-Bench accuracy from 11.2 to 14.6 without further training. That points to reusable mechanisms, not a fit to one model.

Why it matters

Agent progress increasingly comes from the harness around a frozen model, not only from new weights. If that harness can self-improve without overfitting its tests, teams get systems that travel better and cost less at runtime. Code is on GitHub under google-research/rrsi; the project page is regularized-rsi.com.

Dany's take

Self-improvement that only wins on the scoreboard you trained on is a trap. RRSI is a practical reminder: regularize the loop, keep the harness open, and measure transfer. That is how agent stacks stay useful when the next benchmark shows up.