Google DeepMind paper trains an agent to rewrite its own harness (runtimewire.com)

Google DeepMind has introduced Evo2, a groundbreaking framework that enables AI agents to autonomously enhance their software harnesses while simultaneously refining the policy that governs those edits. This innovative approach treats modifications and their evaluations as training data, allowing for a seamless upgrade of the agent's workflow without altering the underlying task executor. The paper, published on arXiv on October 9th, demonstrates significant improvements in coding and memory performance, achieving a 48.3% resolved rate on the SWE-bench Verified benchmarks. The significance of Evo2 lies in its potential to revolutionize the development of AI agents by streamlining the process of improving their operational frameworks. By enabling agents to effectively rewrite and evaluate their harnesses, DeepMind's research could lead to more efficient AI applications, reducing the need for manual intervention and oversight. This self-improvement capability not only enhances performance but also fosters a new paradigm in AI development, where agents can become increasingly autonomous and adaptable, pushing the boundaries of what's possible in AI/ML.