Google researchers introduced Harness², a plug-and-play framework for recursively improving AI agents’ execution harnesses, in a paper published on October 9. The authors say an agent’s capabilities depend on both its underlying model and the harness that runs it. Recursive self-improvement (automatic revision of a system using its own results) faces two open-world challenges: tasks may be unseen when a harness is designed, leaving no ground-truth verifier to guide edits at run time; and improvements may be too task-specific to work in unfamiliar scenarios.

Harness² uses two connected levels of recursion. At the task level, a contrastive Propose-Probe-Compose step refines the harness for the current task without ground-truth verification; this process can run in parallel or sequentially. At the harness level, the framework learns from those edits to update a continuing-improvement harness, helping the agent adapt to later tasks.

Across three professional benchmarks and four model-harness configurations, one step of Harness² consistently improved GLM-5 and Gemini 3.5 Flash over their base harnesses. The reported gains reached 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. The paper also reports further gains as the task-level recursion budget increases.

Google researchers introduced Harness², a plug-and-play framework for recursively improving AI agents’ execution harnesses, in a paper published on October 9. The authors say an agent’s capabilities depend on both its underlying model and the harness that runs it. Recursive self-improvement (automatic revision of a system using its own results) faces two open-world challenges: tasks may be unseen when a harness is designed, leaving no ground-truth verifier to guide edits at run time; and improvements may be too task-specific to work in unfamiliar scenarios.

Harness² uses two connected levels of recursion. At the task level, a contrastive Propose-Probe-Compose step refines the harness for the current task without ground-truth verification; this process can run in parallel or sequentially. At the harness level, the framework learns from those edits to update a continuing-improvement harness, helping the agent adapt to later tasks.

Across three professional benchmarks and four model-harness configurations, one step of Harness² consistently improved GLM-5 and Gemini 3.5 Flash over their base harnesses. The reported gains reached 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. The paper also reports further gains as the task-level recursion budget increases.