Google's Harness² lets an AI agent tune its own harness Google

Google Research has released Harness², an open-source framework that lets an AI agent improve its own harness for a new task without retraining the model. The harness is everything around the model: system prompt, tools, memory, skills, subagents and more. The code is on GitHub under the Apache 2.0 license, with a paper on arXiv.

The idea: an agent's results depend on the model and on the setup around it. Harness² tunes that setup for each task while the model weights stay fixed.

How it works

• Propose, probe, compose. The agent proposes edits to its harness, tests several variants on the same task, compares how the runs behaved, and builds a refined harness from those differences. It needs no ground-truth answer.
• It learns across tasks. After each task, Harness² keeps the adopted harness and the lessons the runs supported. The next round of tuning starts from that experience.
• Parallel or sequential. It can test many candidate harnesses at once, or refine one harness step by step.
• Eight editable parts: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.
• Executors: OpenCode (default) and Codex are supported out of the box.

Results

Google tested it with GLM-5 and Gemini 3.5 Flash. A single Harness² step improved both over their base harnesses:

More recursion steps brought further gains. The figures are Google's own.

Try it

• Needs Linux, Python 3.12 and uv. A local demo runs without any model credentials.
• Benchmark runs use real models (set up through Vertex AI by default) and cost API money, so start with one task.
• Google says it is "not an officially supported Google product".

Sources

• GitHub: google-research/harness2
• Paper: Harness²: Recursive Agent Harnessing for an Open World (arXiv)
• Project site