Google open-sources Harness², which lets agents improve without touching model weights

Google Cloud AI Research and researchers at Northwestern University have open-sourced Harness², a recursive agent-harnessing framework released at version 0.1.0 under the Apache-2.0 licence. It is not a model: it trains nothing and changes no weights.

A harness is everything around the model — system prompt, tools, memory, skills and orchestration. Harness² exposes eight editable components, each mounted as a file: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugin middleware.

The framework runs two levels of recursion. At the task level it uses a contrastive propose, probe and compose loop: run the task once on the base harness, propose two candidate edits — one aimed at general agent failures such as planning and tool use, one at domain knowledge — run both on the same task, then keep only edits backed by observed behaviour.

The second level happens after each task. A reflector stores the adopted harness together with an EXPERIENCE.md file of lessons, which guides later proposals, while every new task still starts from the base harness. A budget k scales the search, either through parallel probes or sequential refinement steps.

The team evaluated on WorkBuddy-Bench, JobBench and Harvey's LAB, using Gemini 3.5 Flash and GLM-5 under OpenCode and Codex, with Claude Opus 4.8 as the default editor. Reported numbers are means over three runs.

At k=1, one refinement step improved all four model-and-harness setups on all three benchmarks. The largest gain was GLM-5 under OpenCode on WorkBuddy-Bench, from 64.2 to 75.7, a rise of 11.5 points; the smallest was GLM-5 under Codex on JobBench, from 23.6 to 24.8. Weaker executors gained more.

Parallel recursion at k=5 pushed Gemini 3.5 Flash under OpenCode to 80.8 on WorkBuddy-Bench. The cost is extra runs per task, and the authors advise sandboxing. The README targets Linux with Python 3.12 and uv, with model calls routed through Vertex AI.

The point is that self-improvement moves from training to use: the model stays the same, yet the agent can get better as tasks accumulate. For enterprises that means a new lever after deployment, though whether the lessons library transfers across domains — and whether the extra compute pays for itself — remains to be seen.