Evo2 proposes co-evolving an AI agent's harness and model weights

A post describing a Google DeepMind paper says Evo2 uses one policy to propose changes to an agent's harness and assess their effects. The edits and outcomes then train both roles through supervised fine-tuning followed by interleaved GRPO. The post reports scores of 48.3% on SWE-bench Verified and 82.1% on Polyglot, against 41.0% and 62.6% for the strongest baselines. It also says harness changes alone raised a GPT-5-mini agent's SWE-bench Verified score from 24.4% to 48.3%.