RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
RRSI regularizes automated harness editing to curb benchmark overfitting, improving held-out tasks while using fewer policy tokens than unregularized evolution.
RRSI constrains recursive harness editing with an annealed proposal budget, exploration history, a critic and a pruner. Across **eight benchmarks**, it reports gains up to **14.1 points** on the evolution split and **4.7 points** on five out-of-distribution benchmarks.
When optimizing an agent harness, evaluate reusable mechanisms on held-out tasks and penalize bundled, costly or negligible edits. The reported harness also used **30% fewer policy tokens** than unregularized evolution, making efficiency part of selection rather than an afterthought.
RRSI constrains recursive harness editing with an annealed proposal budget, exploration history, a critic and a pruner. Across **eight benchmarks**, it reports gains up to **14.1 points** on the evolution split and **4.7 points** on five out-of-distribution benchmarks. When optimizing an agent harness, evaluate reusable mechanisms on held-out tasks and penalize bundled, costly or negligible edits. The reported harness also used **30% fewer policy tokens** than unregularized evolution, making efficiency part of selection rather than an afterthought. The gains are maxima rather than aggregate results, and in-distribution improvement remains much larger than the held-out gain. Performance across the individual coding, workspace and design benchmarks is not provided in the supplied material.
This supplies measured evidence that self-editing harnesses benefit from explicit controls on proposal volume, reuse and cost, rather than unconstrained evolution. Held-out gains support selecting reusable mechanisms, while the much smaller out-of-distribution improvement cautions that benchmark optimization still transfers only partially and does not establish maintainability in production code.