Google's RRSI Open-Sources a Framework That Evolves AI Agents Without Touching Model Weights
Google Research open-sourced a framework that evolves an LLM agent's prompts, tools, control flow, and memory around a frozen model without overfitting.
- Google Research released RRSI, a framework that evolves LLM agent harnesses around a frozen model.
- Regularizes the search trajectory with annealed edit budgets, history-aware proposals, and a leakage-screening critic.
- Every candidate lives in its own git worktree; accepting a winner fast-forwards the branch.
- One loop drives three instances: Terminal-Bench coding, Harvey LAB workspace, and EngDesign engineering.
- Gains transfer out of distribution: +4.3 on Frontier-Eng, +4.7 on JobBench, +1.8 on SWE-bench Verified.
- Apache 2.0, Python 3.10+, works with Claude Opus 4.8 or Gemini 3.5 Flash as the frozen policy.
Google Research has released RRSI source code, an Apache 2.0 framework that automatically improves the harness around a language model whose weights remain fixed. Short for Regularized Recursive Self-Improvement of Agent Harnesses, RRSI searches across prompts, tools, memory, context management and control flow while applying constraints designed to reduce benchmark overfitting.
In an LLM agent, the harness determines how the model receives instructions, calls tools, retains context and coordinates sub-agents. Optimizing those components against a fixed task set can produce large gains on familiar examples while yielding smaller gains on new benchmarks. RRSI addresses that gap by regularizing the search process, meaning it constrains how candidate changes are proposed and accepted.
Edits face two gates
RRSI divides its safeguards between proposal generation and candidate selection. The proposal controls preserve a broad editing surface while discouraging repeated or overly complex experiments; the selection controls reject leakage, noise and unjustified inference cost.
| Stage | Mechanism | Purpose |
|---|---|---|
| Proposal | Annealed edit budget | Limits how many independent changes a candidate can bundle, with the budget narrowing over time. |
| Proposal | Full edit history | Shows the proposer previous hypotheses, results and rejections so failed ideas are less likely to recur. |
| Proposal | Stagnation redirection | Steers stalled searches toward harness components that have received little or no attention. |
| Selection | Leakage critic | Rejects suite-specific logic before evaluation, including patterns identified by the domain adapter. |
| Selection | Noise-adjusted floor | Blocks apparent gains that fall within measured evaluation variance. |
| Selection | Token-cost rule | Requires measured improvements to justify additional inference tokens. |
| Selection | Component pruning |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
