Google Research releases Harness² framework for agent self-improvement
The implication is that harness scaffolding, not model weights, is a lever for gains on unseen tasks, with a frozen model still improving when its surrounding code is edited. Developers building agent systems can apply the editor-composer-reflector pattern to their own pipelines. Code and the arXiv paper (2610.14224) are public, so teams can reproduce the benchmark setup directly.
Google Research released Harness², a framework that lets agents improve the harness around a frozen model. The pipeline runs a base harness to record baseline trajectories, then an editor agent drafts two candidate edits such as skills and hooks, a composer merges edits that fix failures, and a reflector stores verified lessons in a library. On unseen tasks, Gemini 3.5 Flash gained +8.4 on WorkBuddy-Bench and +7.5 on JobBench, improving on all 3 benchmarks tested. Skills and scripts account for roughly 50% of edits, and dropping skills costs 5.3 points.
Can Agents improve their own harness on unseen tasks? The @GoogleResearch team released Harness², a framework that allows improve an harness around a frozen model. How it works: 1️⃣ Agent runs the base harness on each task to record a baseline trajectory 2️⃣ Editor agent inspects the trajectory, drafts 2 candidate edits (SI, Skills, hooks,….), and runs both. 3️⃣ Composer agent diffs the new trajectory against the baseline, merges edits that fixed failures (within a cap), and runs the final harness. 4️⃣ Reflector agent logs verified lessons into file and saves the winning components to a library for future tasks to pull on demand. Results: - 💡 Gemini 3.5 Flash +8.4 on WorkBuddy-Bench and +7.5 on JobBench. - 🛠️ Skills and scripts drive ~50% of edits; dropping skills costs 5.3 pts. - 📚 Gemini 3.5 Flash improves on all 3 benchmarks when editing its own harness. Pape arxiv.org/abs/2610.14224 GI0 Cod github.com/google-researc… M9s 💬 13 🔄 2 ❤️ 24 👀 1391 📊 15 ⚡
