The Pulse
Google Researchers Add Guardrails to Agent Self-Improvement
A paper released by Google Cloud AI Research collaborators on September 21 describes a method for limiting how AI agents revise their operating instructions and how those revisions are selected. Its experiments report that the method improv
AI.info Team · Written with AI ·
Harness-evolving agents can post better benchmark scores while learning changes that fail to travel beyond the tasks used to tune them. A paper released by Google Cloud AI Research collaborators on September 21 proposes a countermeasure: constrain how agents revise their operating instructions and how those revisions earn a place in the system.
The method, called Regularized Recursive Self-Improvement of Agent Harnesses (RRSI), targets the “harness” around a frozen language model: its prompts, control flow, tools, memory and context management. The authors test it across coding, agentic workspace and engineering-design tasks, reporting gains on held-out benchmarks. Their experiments also show the tension behind the work: unrestricted evolution can score higher on its target set while transferring less reliably.
RRSI limits what each round can change
Instead of allowing a search process to pile many edits into each proposed update, RRSI gradually tightens a budget on the number of changes a candidate can bundle. It tracks which edits were tried, what each was meant to fix, and whether it helped. When progress stalls, the system can direct proposals toward components it has not yet explored.
A second set of checks governs which candidates survive. A critic screens for benchmark-specific logic, such as edits that encode task names or answers. A noise-adjusted performance floor is intended to prevent a lucky evaluation from becoming a permanent change; cost checks limit added inference tokens, and a pruning rule removes components that stop helping.
“Across eight benchmarks spanning coding, agentic workspace and engineering design tasks, RRSI gains up to 14.1 points on the split it evolves against and up to 4.7 points on the five out-of-distribution benchmarks, while producing a harness that runs on 30% fewer policy tokens than the unregularized evolution.”
Peng Xia and co-authors, authors of the paper, write.
The largest gain depends on the model and test
The 14.1-point result came from a coding experiment using Gemini 3.5 Flash as the frozen policy: Terminal-Bench 2.1 rose from 64.6 to 78.7. On the unseen SWE-bench Verified set, the same evolved harness gained 2.2 points, moving from 76.8 to 79.0. In a separate run with Claude Opus 4.8, Terminal-Bench rose by 6 points, while SWE-bench Verified improved by 1.8.
For the paper’s main cross-domain results, the researchers used Claude Opus 4.8 as the frozen policy. The evolving system improved every held-out split they tested: JobBench rose 4.7 points, GDPval 3.5, APEX-Agents 3.7, Harvey LAB’s held-out tasks 2.3, SWE-bench Verified 1.8, and Frontier-Eng 4.3 medal points. The tasks cover workplace assignments, software fixes and engineering designs; the authors say the engineering benchmarks use deterministic simulators or test benches rather than model-based judges.
Higher training scores did not mean broader gains
The clearest comparison comes from the agentic-workspace ablation. RRSI scored 90.5 on the Harvey LAB evolve set and averaged 43.6 across JobBench, GDPval and APEX-Agents. Unregularized evolution scored higher on Harvey LAB, at 92.8, but its out-of-distribution average was 40.3—close to the unevolved harness’s 39.7.
The cost comparison complicates the result. In that same experiment, RRSI used 2.42 million policy tokens per trial, compared with 3.80 million for unregularized evolution. But the unevolved harness used 1.56 million. RRSI’s controls reduced the cost of evolution; they did not make the evolved system cheaper than leaving the harness untouched.
The paper stops at the harness
RRSI changes the instructions, tools and other scaffolding around a model, not the model’s weights. The authors say their study does not establish how the method would work when evolution updates those weights. They also rely on a finite set of tasks for evolution and on regularization settings that may affect results, leaving the quality of the feedback and the size of the search budget as limits.
The researchers have published the implementation in the Google Research RRSI repository under the Apache 2.0 license. The repository states that the project is not an officially supported Google product.