Harness² Explained: Self-Improving Agents Without Training

Harness² is a research framework from Google Cloud AI Research and Northwestern University that lets an AI agent rewrite its own harness (its prompts, skills, scripts, subagents and control-loop hooks) for every new task, without retraining the model. In this video I explain how it improves a harness with no answer key (Propose, Probe, Compose), why it has two levels of recursion with an experience library, what the results show against Claude Code and Codex, and where the approach breaks. Paper: https://arxiv.org/abs/2610.14224 Code: https://github.com/google-research/harness2 Project page: https://harness2.github.io What is an Agent Harness?: https://youtu.be/nWzXyjXCoCE Claude Can Now Build Its Own Harness: https://youtu.be/l5rae4LMKBc DeepSeek Harness: https://youtu.be/NPO2CwHnfmI What Are Claude Code Mods?: https://youtube.com/shorts/eN77oHqAiUU 0:00 The Flaw of One-Size-Fits-All AI Harnesses 1:02 Harness Squared: Agents That Evolve Themselves 2:16 Four Pillars of a Smarter Agent Architecture 4:52 Propose, Probe, and Compose: The Validation Loop 7:47 Real-World Takeaways and Agentic Risks