The Pulse
RRSI Paper Describes a Method for Improving Agent Harnesses
The arXiv abstract reports that RRSI improves performance on both the split used for evolution and five out-of-distribution benchmarks while reducing policy-token use. No attributable quotation appears on the arXiv abstract page; it lists t
AI.info Team ·
A paper published on arXiv describes a method for improving an AI agent’s harness—the prompts, control flow, tooling, memory and context management surrounding a frozen language model. Called Regularized Recursive Self-Improvement of Agent Harnesses, or RRSI, the method repeatedly proposes and selects edits to that surrounding system rather than retraining the underlying model.
The paper frames the approach as recursive self-improvement at the agent-system level. The backbone model remains frozen while the harness changes through an automated process that evaluates candidate edits. The authors say existing approaches can overfit to the tasks used during improvement, producing large gains on those tasks that shrink or disappear when the system is tested elsewhere.
Results on eight benchmarks
RRSI is evaluated across eight benchmarks spanning coding, agentic workspace tasks and engineering design. The paper reports gains of up to 14.1 points on the split used to evolve the harness. On five out-of-distribution benchmarks, the method produces gains of up to 4.7 points.
The evolved harness also runs on 30% fewer policy tokens than an unregularized evolution process. In the paper’s description, the result is intended to show that constraints on the search can favor changes that continue to work beyond the tasks used to generate them, rather than changes that simply memorize benchmark-specific patterns.
Constraints on candidate edits
RRSI limits the number of edits that a candidate can bundle through a temporally annealed budget. The proposal process can use the evolution history to encourage unexplored trajectories, reducing the chance that it repeatedly follows the same unsuccessful path.
The selection process adds two components: a critic and a pruner. According to the abstract, the critic screens proposals that appear specific to the benchmark. The pruner removes changes that are too small, too expensive or no longer useful. Together, these restrictions are intended to favor reusable agent mechanisms over benchmark-specific changes or noise.
The paper describes the harness as a broad software layer around the model. Its editable elements can include prompts, control flow, tools, memory and context management. This allows the method to search for changes in how an agent is organized and operated while keeping the underlying language model fixed.
What the paper claims
The central claim is narrower than an agent independently redesigning an entire AI system. RRSI is a controlled procedure for proposing, testing and selecting changes to a defined harness. Its reported results concern benchmark performance and policy-token use under that procedure.
The authors say the method addresses a specific weakness in recursive harness evolution: improvements on the tasks used for development may not transfer to new tasks. By constraining both candidate proposals and the selection process, RRSI aims to identify changes that remain useful outside the split used for evolution.
The paper was submitted to arXiv on September 21, 2026. Its abstract says that code is available through a linked project resource.