The report says Google open-sourced Harness² to automatically optimize the AI Agent toolchain

AT A GLANCE

The report says Google Research has open-sourced Harness², which can automatically improve the AI Agent toolchain without retraining the model.

Article

Google open-sources Harness²: enabling AI to improve its own toolchain,GLM-5 benchmark scores increased by 11.5 points

Beating AI breaking news: Google Research has open-sourced Harness², enabling AI Agent to automatically modify system prompts, memory,Skills, tool rules, and even scripts and plugins without retraining the model. The project uses the Apache 2.0 protocol and supports OpenCode and Codex.

Its core is to let Agents conduct their own trial and error. When encountering a new task, the system first proposes several improvement plans, then lets different versions of the Agent execute them separately, compares their performance, and finally combines a more suitable plan.

Even without a standard answer, it can identify directions for improvement based on the actual execution process. After a task is completed, the system also remembers which modifications were effective and why, and continues to refer to them when facing similar problems next time rather than starting from scratch each time.

Google tested using GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench,GLM-5 paired with OpenCode, after just one round of optimization, increased its average score from 64.2 points to 75.7 points, exceeding the unoptimized Gemini 3.5 Flash's 72.4 points. In legal-task testing,GLM-5's pass rate on certain scoring criteria also rose from 59.6%to 71.4%.

However, in the experiment, Claude Opus 4.8 was responsible by default for proposing and implementing improvements, while GLM-5 and Gemini were primarily responsible for executing tasks; they did not achieve evolution entirely on their own.

Domestic developer Joseph Zhou also demonstrated an experiment integrating Harness² with Jev 1.14, claiming that performance on visual tasks continued to improve.

Original link https://m.theblockbeats.info/flash/372535

Key points

Beating AI reports that Google Research has open-sourced Harness².
The report says Harness² can modify system prompts, memory,Skills, tool rules, scripts, and plugins without retraining the model.
The project is said to use the Apache 2.0 protocol and support OpenCode and Codex.
The report says the system proposes multiple improvement plans, lets different versions of the Agent execute them and compare performance, then combines the plans.
The report says the system can record effective modifications and their reasons, and continue to refer to them for similar problems.
The article states that on WorkBuddy-Bench,GLM-5 paired with OpenCode increased its average score from 64.2 points to 75.7 points after one round of optimization; the unoptimized Gemini 3.5 Flash scored 72.4 points.
The article states that in legal-task testing,GLM-5's pass rate on certain scoring criteria rose from 59.6%to 71.4%。
In the experiment, Claude Opus 4.8 proposed and implemented improvements, while GLM-5 and Gemini were primarily responsible for executing tasks.
Joseph Zhou demonstrated an experiment integrating Harness² with Jev 1.14 and claimed that performance on visual tasks continued to improve.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Harness² is described as a toolset that enables AI Agent to conduct trial and error using multiple approaches, compare results, and retain effective modifications. Here, “self-improvement” primarily refers to adjusting prompts, memory, and tool configurations, rather than retraining the model; moreover, the report explicitly states that the improvement work in the experiment was handled by Claude Opus 4.8 by default.

Why it matters to readers

If the functionality and test results described in the article can be reproduced, Harness² demonstrates an optimization approach that improves Agent configurations and toolchains rather than retraining models.

Beginners should distinguish between “optimizing an Agent toolchain” and “training a model”: the performance improvements cited in the article stem from modifications to system prompts, memory, tool rules, and similar elements, and do not mean the model parameters themselves changed.

Risks and unknowns

The evidence consists only of media reporting, without an official Google Research announcement or code repository; the project's affiliation, open-source status, and Apache 2.0 license have not been cross-verified.
The complete methodology, samples, baselines, and independent reproduction results for the WorkBuddy-Bench and legal-task tests were not provided.
The actual degree of compatibility with OpenCode and Codex cannot yet be confirmed.
Joseph Zhou's claim of ongoing improvements to visual tasks does not provide specific data or validation methods.

The report says Google open-sourced Harness² to automatically optimize the AI Agent toolchain

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center