In early October 2026, coverage by The Decoder and the project site regularized-rsi.com highlighted a Google Cloud AI Research paper — with co-authors from UNC-Chapel Hill, Stanford and Washington University in St. Louis — titled RRSI: Regularized Recursive Self-Improvement of Agent Harnesses (arXiv:2609.24972).
The core claim: an LLM agent’s capability is largely magnified by its harness — prompts, control flow, tooling, memory and context management — around a frozen backbone model. Recent methods iteratively propose and select component-wise edits, a practical form of recursive self-improvement at the agent-system level. That evolution, however, often overfits by memorising the training tasks.
RRSI builds regularisation into candidate proposal and selection. The proposer runs with a temporally annealed budget: it limits how many edits a candidate may bundle and favours unexplored trajectories based on evolution history. Early rounds allow larger rewrites; later rounds permit only small changes clearly tied to a measured result.
The selector combines a critic and a pruner. The critic screens out benchmark-specific proposals (task names, hard-coded solutions or other tricks). The pruner removes changes that are too small, too expensive or no longer useful. Together these constraints favour reusable agent mechanisms over evolve-set shortcuts.
The team tested RRSI on eight benchmarks spanning coding, agentic office work and engineering design. The underlying model, Claude Opus 4.8, stayed frozen. They compared RRSI with the unmodified baseline harness and four recent optimisation methods.
According to the paper, RRSI gains up to 14.1 points on the split it evolves against and up to 4.7 points on five out-of-distribution benchmarks, while producing a harness that runs on about 30% fewer policy tokens than unregularised evolution. Overall performance never fell below the baseline on any unseen benchmark — the typical failure mode of a harness that memorised its tests.
A transfer result: a coding harness optimised with Gemini 3.5 Flash raised the accuracy of the weaker Gemini 3.1 Flash Lite from 11.2 to 14.6 points without modifying the weak model. The mechanisms discovered do not depend on the capability of the model used to find them.
The contrast with hand-designed harnesses is stark. Cited ARC-AGI-3 tests showed a purpose-built harness scoring 97.1% in a familiar environment and 0% in an unfamiliar one with Opus 4.6. Nvidia recently presented SoL-Pi, a related method in which a research agent automatically rebuilds coding-agent harnesses and cuts token use by up to 49% without a noticeable drop.
Code is at github.com/google-research/rrsi and the project page at regularized-rsi.com. Listed authors include Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhuang, Yoonho Lee, Chengsong Huang, Han Yu, Zhongying CuiZhu, Yifei Ming, Huaxiu Yao, Burak Gokturk, Tomas Pfister and Chen-Yu Lee.
For product labs, RRSI arrives as the industry debates RSI at the weight level versus the system level. Regularising the latter is cheaper and more auditable than retraining the backbone, and it reduces the risk of reporting benchmarks inflated by memorisation.
Honest limits: gains depend on domain, judge and evolution window; 14.1 points is a reported ceiling, not a guarantee. The paper uses Claude Opus 4.8 as the policy — an expensive frontier model — to discover mechanisms that can later transfer to cheaper models.
In the public debate about “pacing the frontier,” RRSI offers a technical nuance: agentic behaviour can improve without releasing a more capable model, yet harness self-improvement can still outrun auditors if components are not versioned.
Neural Pulse recommends evaluation teams adopt mandatory held-out suites, filter proposals that name the benchmark, and publish final harness token cost alongside the score — metrics RRSI foregrounds.
The work connects to October 2026 anxiety about self-improving systems: here the improvement is of scaffolding, not weights, but the propose–select loop remains a practical RSI form that regulators and red teams must instrument.
In short, Google and academic partners deliver a concrete method so agent harnesses stop “learning the exam,” generalise better, spend fewer tokens — and ship open code to reproduce.
