google-research/harness2 is public research code that lets an agent iteratively improve its task-execution tools, instructions, and memory configuration while keeping the model fixed. Here, a harness means an environment that bundles a system prompt, tools, memory, and skills so the model can perform tasks. It is not an approach that separately changes model weights. The repository carries no information about a latest release.
This framework compares the results of current task executions to produce the next configuration, and accumulates grounded improvements into later tasks as well. The core is a cyclical structure that reflects per-task results into the next setup. So the focus of the improvement is not the model itself but the execution environment and the instructions surrounding it. The repository also includes run, score, and report tools for several evaluation suites covering professional and legal work, along with setup instructions. It is written in Python and released under the Apache-2.0 license. The fact that evaluation tools are provided does not mean that the improvements have been demonstrated across multiple kinds of work.
The performance figures are test results presented by the repository and cannot be treated as independently confirmed. From this material alone, it is difficult to judge whether the presented improvements reproduce under external conditions. So the core of harness2 lies in a research framework that repeatedly tunes the harness without changing the model and accumulates the results; its performance effect should be read within the scope of the repository's own tests until independently verified.