Google Lets Agents Rewrite Themselves

🔥 What happened Google Cloud AI Research, with UNC, Stanford and WashU, dropped RRSI: an LLM agent that rewrites its own harness — prompts, tools, memory, control flow, sub-agents. Model weights stay frozen. Apache 2.0, Python 3.10+, any LiteLLM model string. 💡 Why it matters Instead of training weights, RRSI regularizes the agent shell itself to stop overfitting. On SWE-bench Verified — never used for selection — it climbs 82.0% to 83.8%. With Gemini 3.5 Flash, Terminal-Bench 2.1 jumps from 64.6% to 78.7%. ⚡ Our take The real 2025 lever isn't smarter models — it's smarter scaffolding. Teams still tuning weights will get lapped by anyone regularizing their agent's self-improvement loop.