Google Harness²: AI Agents That Fix Their Own Setup (I Tested It)

Google Research just open-sourced Harness², a framework that lets an AI agent rewrite its own harness (the system prompt, guardrails, memory and skill files around the model) while it is doing the job, with no retraining. I ran it myself on a real legal job with Gemini 3.5 Flash: 20 of 55 checks passed on the first try, 22 of 55 after the agent fixed its own setup. In this video: → What a "harness" is, and what it means for one to improve itself → Where to get the code and what you need to run it → How it works, step by step, using Google's own animation → My own run: the playbook the agent wrote for itself and the one line it added to its system prompt → My +2 next to Google's numbers (their best case went from 44 to 94) → Five real jobs from the paper and the rule the agent wrote for each The demo: one task from the LAB legal test set (reviewing a draft procedural order in an ICC arbitration), with Gemini 3.5 Flash doing the job, the fixing and the grading. The whole run cost about $11. Harness² is research code from Google Cloud AI Research and Northwestern University, not an official Google product. Chapters 0:00 What if the harness improves itself? 0:43 Quick look at my result 1:23 Harness² explained (the new-hire analogy) 1:53 Where to get it 2:20 How it works, step by step 2:56 My test run: what the agent rewrote 3:36 What happened in the run 4:03 Which checks changed 4:24 My +2 in context 4:56 Five real jobs from the paper 5:45 Wrap-up Links Paper (arXiv): https://arxiv.org/abs/2610.14224 Code (GitHub): https://github.com/google-research/harness2 Project page: https://harness2.github.io/ All opinions are my own and do not belong to my employer. #Harness2 #AIAgents #GoogleAI