Harness²

What it is

An AI agent's performance depends heavily not only on the underlying model but also on the execution environment wrapped around it, that is, the harness. According to reports, Harness² focuses on this and modifies only the harness, without changing the model's weights (the internal values set by training). Here, the harness refers to eight components attached as files: the system prompt, guidelines and safety rules, memory, tool-usage instructions, subagents, skills, scripts, and plugins/middleware. Google reportedly posted the research to the online archive arXiv and released the code as open source on GitHub.

How it works

Improvement repeats in two levels. At the first, task level, the agent attempts a task with the base harness, and the improvement system then proposes two revisions. One targets problems common across many tasks, such as planning or tool use; the other fixes knowledge or procedures for a specific domain. Both revisions are applied to the same task, their execution traces are compared (Propose-Probe-Compose), and the changes that showed an effect are selected. At the second, harness level, the experience gained this way is written to an improvement library called LESSONS and consulted in later tasks. Each new task starts again from the base harness rather than the most recently revised setup, a design meant to reduce cases where a fix tailored to one task causes side effects in others.

Why it matters

Common ways to improve model performance are retraining the model or tuning it against evaluation data with fixed correct answers. In real work, however, unfamiliar tasks keep appearing, and often there is no criterion for automatically judging whether a result is correct. Harness² differs in that it decides the direction of improvement by comparing two execution results against each other, without a verifier that supplies the correct answer. According to reports, the researchers tested it with GLM-5 and Gemini 3.5 Flash on professional-work benchmarks WorkBuddy-Bench, JobBench and Harvey's LAB, and after a single round of improvement all four model–harness combinations scored higher on all three benchmarks. The largest gain was GLM-5 with OpenCode on WorkBuddy-Bench, which reportedly rose 11.5 points, from 64.2 to 75.7. On JobBench the gain was reportedly up to 4.5 points, and on LAB, a legal-work evaluation, the per-item pass rate rose by up to 11.8 percentage points.

Things to keep in mind

The information above is based on press reports, and the performance figures are results stated by the researchers. Gains vary by combination: GLM-5 with Codex rose only 1.2 points on JobBench, from 23.6 to 24.8. The researchers said that spending more budget on improvement to explore revisions multiple times could raise performance further. Because the model itself is not changed, it does not gain abilities it lacked; it is more accurate to see this as a way of drawing out abilities the model already has. Readers should distinguish the general concept of a "harness" from "Harness²," the name of this specific framework.

Example

On this site, it was covered in the news item "Google releases Harness², an open-source framework that improves the agent execution environment on its own while leaving the model unchanged (reported)." For example, imagine a coding agent assigned to internal document work. Instead of retraining the model, you take the process the agent actually followed and run two versions side by side: one with an added general instruction such as "check the tools first," and one with an added domain-specific instruction such as "organize documents in this domain in this order." If you adopt the changes from the better version and record the lessons learned, the next task starts from the base setup but consults those lessons when proposing new revisions.

Where it came up on this site

• briefing Google releases 'Harness²', an open-source framework that improves the agent execution environment on its own while leaving the model unchanged (reported) 2026-10-19

Related terms

Harness AI agent System prompt Skills Agent memory Subagent

References

• Google releases 'Harness²', a self-improvement framework that raises AI performance on its own (AI Times)
• Google Open-Sources Harness2, a Recursive Agent Harnessing Framework (MarkTechPost)
• Google releases Harness2, AI agents that improve themselves (Pasquale Pillitteri)
• Google Open-Sources Harness² Framework: No Model Weight Changes Needed, Agents Boost Scores by Up to 11.5 Points by Rewriting Their Own "Shell" (BigGo Finance)