Google DeepMind paper trains an agent to rewrite its own harness

Google DeepMind researchers introduce Evo2, a framework that trains an AI agent to rewrite and maintain the software harness surrounding its model while simultaneously training the policy that proposes and evaluates those edits. Evo2 pairs concrete code edits with performance feedback and treats both the edits and the evaluation traces as training data, so the system co-evolves harness code and model weights without altering the underlying task executor. The approach explicitly separates a task executor from a maintainable harness and trains an editing policy to propose improvements plus an evaluator to score their impact, creating an iterative loop where harness changes become part of the agent’s learning signal.

Evo2 demonstrates measurable gains on coding and memory benchmarks, including a 48.3% resolved rate on SWE-bench Verified, indicating substantial improvements in problem resolution after harness-driven adaptation. By logging harness changes and using those logs as supervised signals, Evo2 improves workflow robustness and the policy that maintains that workflow, delivering a practical path to self-improving agents that refine their runtime scaffolding. The framework positions harness editing as a scalable mechanism for continual improvement, showing that operational code and evaluation traces are valuable training material alongside traditional model updates.