Google Open-Sources Harness² , a Framework That Improves AI Agents Without Retraining Models

Google Research has released an open-source framework called Harness² that helps AI agents improve their performance by adjusting the systems around them, not by changing the underlying model itself.

The idea is straightforward: an AI agent's success depends not just on the model running it, but on everything surrounding it. That includes the instructions given to the model, the tools it can use, how it remembers past interactions, and how it handles decisions. Harness² treats all of that as adjustable components that can be refined over time.

"Your agent may get better without a better model," the research team writes.

The framework is version 0.1.0 and runs on Linux with Python 3.12 or 3.13. It does not train any model weights. Instead, it works with eight editable components, each stored as a separate file: system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins ¹.

Harness² uses a two-level recursive process to make improvements. First, when a task comes in, the agent runs it once using its current setup. Then an editor proposes two possible changes: one aimed at fixing general problems like planning or tool use, and another targeting knowledge specific to the task's domain. Both versions are tested on the same task, and the system keeps only the changes that actually produced better results.

After each task, a reflector stores what was learned in a file called EXPERIENCE.md. This creates a growing library of lessons that guide future improvements ¹.

The researchers tested Harness² on three benchmarks: WorkBuddy-Bench, JobBench, and Harvey's LAB. They used Gemini 3.5 Flash and GLM-5 as the base models, along with OpenCode and Codex as execution environments.

The results showed meaningful gains after just one refinement step. GLM-5 under OpenCode improved from 64.2 to 75.7 on WorkBuddy-Bench, a jump of 11.5 points. Even the weakest performers saw improvement across all three benchmarks ¹.

When the team ran more refinements in parallel mode with k=5, Gemini under OpenCode reached 80.8 on WorkBuddy-Bench. Three of the four tested setups also required fewer steps per task after harnessing ¹.

Compared to other methods, Harness² stands out because it learns across tasks rather than optimizing for a single one. A previous approach called Meta-Harness searches for code offline, while ReasoningBank only adjusts retrieved memory with a fixed harness ¹.

Installing Harness² requires Linux, Python 3.12, and a package tool called uv. A local demo can run without model credentials, but full model-backed experiments need Node.js, the Google Cloud CLI, and a Vertex AI project ¹.

The code is available under the Apache-2.0 license, making it freely usable for developers and researchers.

Key facts

• Harness² improves AI agents by refining their surrounding system, not by changing model weights
• One refinement step raised GLM-5 scores by 11.5 points on WorkBuddy-Bench
• The framework uses eight editable components including prompts, tools, memory, and skills
• A lessons library stores experiences to guide future task improvements
• It runs on Linux with Python 3.12+ and is open-sourced under Apache-2.0

Sources

1. Google Open-Sources Harness2, a Recursive Agent Harnessing Framework — marktechpost.com