Google Research has released Harness2, an open-source framework that changes an AI agent’s harness at test time. A harness is the surrounding setup that helps a model carry out a task; the source does not explain exactly which parts Harness2 rewrites or how it does so.
MarkTechPost reports that one refinement step raised GLM-5’s score by 11.5 points on WorkBuddy-Bench, with no updates to the model’s weights. That suggests the reported improvement came from changing the agent setup rather than retraining the underlying model. The article does not give the starting score, the test configuration or the group that ran the benchmark, so the result cannot be independently assessed from the information provided.
What the release tells developers
Harness2 is released under the Apache-2.0 license. That license allows developers to use, modify and share the software subject to its terms, giving research teams and other developers a way to examine the framework rather than relying only on a description of it.
The announcement is narrow: it reports one model, one benchmark and one refinement step. It does not say whether the gain holds for other models or tasks, what computing resources the method needs, or whether the framework is ready for routine use. Those details matter before developers can judge whether Harness2 will improve their own agents.
Why it matters
AI developers can inspect and adapt Harness2 under the Apache-2.0 license, and the reported result suggests agent setup may affect benchmark scores even without retraining a model. But the source gives no test details or evidence that the gain carries over to other models or tasks, so its practical value remains uncertain.
