- Harness² is a framework that leaves the AI model itself unchanged and instead adjusts the settings around the model (system prompts, memory, tool guidance, skills, and so on) for each task, helping agents perform better. The repository is published under the google-research GitHub organization, and the README states that it is not an official Google product.
- During the task phase, it creates revised settings, applies them to the same task, compares the differences in results, and combines the new settings. According to the README, this process runs only on differences between executions, without correct-answer feedback.
- The README reports that in a single-step run across three specialized benchmarks and four combinations of models and settings, GLM-5 and Gemini 3.5 Flash improved over their existing settings. The largest gains are 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points in LAB partial pass rate.
- Running it assumes a Linux and Python 3.12 environment. Benchmarks that use models incur API costs, and a local demo that runs without model credentials is also provided.
O: Developers who hand office work to AI agents may want to split their instructions and tool guides into two versions, run both on the same task, and compare the differences in results.
X: The performance gain figures in the README were reported by the development team and are not independently verified results, so check the model call costs and measurement conditions first.