Inside Google's new self-evolving agent framework

[Sevam]

Google Cloud AI Research and Northwestern University just published a wild paper on Harness² (Harness Squared). Instead of bloating system prompts, this framework lets an agent rewrite its own skills, scripts, and control-loop hooks on the fly using a 3-step Propose, Probe, Compose validation method.
The craziest takeaway? A smaller model inside a self-evolving harness completely outperformed a frontier model running on its default setup. It even has an experience library to carry over optimized hooks to future tasks.