Google Research has open-sourced Harness², a self-improvement framework for AI toolchains that lets AI agents modify their own system prompts, memory, skills, tool rules, scripts, and plugins without retraining the underlying model. Released under the Apache 2.0 license, the project supports OpenCode and Codex.
The framework works by having an agent generate several candidate improvements for a new task, run different agent variants against those options, compare performance, and combine the better-performing changes into a more suitable setup. Google said the system can still search for useful improvements even when no standard answer is available, using the execution process itself as feedback. After each task, it stores which changes worked and why, so those adjustments can be reused on similar tasks later.
In testing, Google used GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 with OpenCode improved from an average score of 64.2 to 75.7 after one optimization round, beating unoptimized Gemini 3.5 Flash at 72.4. In legal-task testing, GLM-5’s pass rate on part of the scoring criteria rose from 59.6% to 71.4%.
Google Research has open-sourced Harness², a self-improvement framework for AI toolchains that allows AI agents to update their system prompts, memory, skills, tool rules, scripts, and plugins without retraining the model itself, according to ChainCatcher.
The project is released under the Apache 2.0 license and supports OpenCode and Codex. Its core method is trial and error by the agent. When facing a new task, the system first proposes several improvement plans. Different agent versions then execute those plans and compare the results, after which the framework combines the more suitable changes.
Google said the system can still look for ways to improve even when there is no standard answer, using the real execution process as the basis for adjustment. After a task is completed, it records which modifications were effective and why, so the same experience can be referenced in similar tasks later.
Test results
Google tested the framework with GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode improved its average score from 64.2 to 75.7 after just one optimization round, higher than the 72.4 scored by unoptimized Gemini 3.5 Flash.
In legal-task testing, GLM-5’s pass rate on part of the scoring criteria increased from 59.6% to 71.4%.
Experiment setup
In the experiments, Claude Opus 4.8 was the default model used to propose and implement improvements. GLM-5 and Gemini mainly handled task execution, meaning the system’s evolution was not achieved entirely on its own.
Developer Joseph Zhou in China also demonstrated an experiment connecting Harness² with Jev 1.14, saying performance on visual tasks kept improving.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
