Google Research has open-sourced Harness², a framework that improves an agent by rewriting the system around it rather than the model weights. A harness covers the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins, all of which Harness² exposes as editable components for a given task. Its task-level recursion runs a Propose–Probe–Compose step: the agent proposes edits, probes their effect on the same task, and composes a refined harness from behavioral contrasts between runs, with no ground-truth answer and no external judge. A second, harnessing-level loop keeps a persistent improvement harness with adopted harnesses and evidence-supported lessons, so the next round starts from what the previous runs proved, in either parallel or sequential mode. Across three professional benchmarks on GLM-5 and Gemini 3.5 Flash, one step lifted results by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB’s partial pass rate, with larger recursion budgets adding further gains; executors supported are OpenCode and Codex.

Related: Stanford + MIT: Meta-Harness Optimization Boosts LLM Performance Up to 6x, Hugging Face Releases Multi-Harness RL for Training Models Inside Coding Agents, Databricks Open-Sources Omnigent, a Meta-Harness for AI Agents

Harness² on GitHub · Paper (arXiv) · Project site