Self-evolving agents stop splitting weights from harness: Evo2 feeds each into the other, resolving 48.3% of SWE-bench Verified

Self-evolving agents have followed two separate routes: one updates model weights from self-generated experience, the other rewrites the harness program that defines how an agent uses prompts, tools, context, and execution feedback. Each route accumulates experience in model parameters or in executable programs, and the two are typically pursued separately, leaving the relationship between weights and harness unused. Evo2 uses harness modifications and their execution outcomes as training signal, and on Polyglot, SWE-bench Verified, and LiveCodeBench the best evolved harness beats the four self-evolution baselines SICA, DGM, HGM, and GEA.

A single policy serves as both Proposer and Reflector: the Proposer writes harness edits, and the Reflector reads execution outcomes and issues a revise, crossover, or keep directive. Harness search is organized as a tree of variants, and the Reflector can pick a strategy from a library that addresses a weakness in the current harness for the Proposer to adapt to that source. Weight updates run through EvoTIDE, an SFT warm start followed by interleaved GRPO over the two roles, trained on trajectories recovered by edit extraction and prefix continuation, with a parent pool that keeps both strong candidates and weaker harnesses needing repair.

The report covers Evo2's self-evolution results on coding and memory tasks: held-out performance on Polyglot, SWE-bench Verified, and LiveCodeBench, six cross-benchmark transfer directions, proprietary-model experiments on a 60-task SWE-bench Verified subset, and memory tasks on LoCoMo and LongMemEval. The task executor stays fixed during co-evolution, so candidate performance measures the effect of changing the harness. It is aimed at researchers and engineers working on agent self-evolution, harness search, and reinforcement-learning training pipelines.

The proprietary-model experiments run on a 60-task SWE-bench Verified subset, with three independent searches per method and model on a small search subset and the mean over seeds reported; baselines evaluate their top-ranked harness while Evo2 reports the better of its two top-ranked harnesses. On memory tasks each search evaluates 1,500 questions drawn from seven training conversations, and LongMemEval is unseen during both evolution and selection. These settings bound how far the current results extend.