Google DeepMind paper trains an agent to rewrite its own harness
Posted on October 9th, Evo2 pairs code edits with performance feedback to train the model proposing and evaluating those edits.
By Ryan Merket · Published · Updated
Primary source: X
Why it matters
Evo2 treats agent harness edits and their evaluation traces as training data, offering a path to improve an agent's workflow and the policy that maintains it without changing the task executor.
Google DeepMind researchers have published Evo2, a framework that trains an AI agent to improve the software harness around its model while also training the policy that edits and evaluates the harness. The paper, posted to arXiv on October 9th, reports gains on coding and memory benchmarks, including a 48.3% resolved rate on SWE-bench Verified.…