#Agent Harness Evolution (1)

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

RRSI regularizes recursive harness evolution while leaving prompts, control flow, tools, memory, context, and subagents editable. Proposal-side regularization anneals edit size, preserves falsified hypotheses as evidence, and redirects stalled searches; selection-side regularization screens leakage, noise, unjustified token growth, and unproductive structure. Across eight coding, workspace, and engineering benchmarks, RRSI improves held-out and out-of-distribution scores and uses about 30% fewer policy tokens than unregularized evolution.