Google open-sources Harness2, letting AI agents rewrite their own harness

Google Research has released Harness2, an open-source framework for AI agents, under the Apache-2.0 license. Announced on 18 October 2026, the tool rewrites the harness surrounding an agent while the agent is being tested, rather than leaving that scaffolding fixed.

According to the published summary, a single refinement step raised the score of the GLM-5 model on WorkBuddy-Bench by 11.5 points. That gain reportedly came without any updates to the model's weights.

Why the harness matters

In agent systems, the harness is the layer wrapped around a model: its instructions, the tools it can call and the control logic that decides how a task unfolds. The same model can perform very differently depending on that setup. Harness2 takes a recursive approach, letting the system redesign this layer instead of relying on engineers to tune it by hand.

Because no retraining is involved, the approach points to a way of extracting more capability from existing models without costly training runs. That could appeal to teams that lack the resources to build or fine-tune their own models.

Open questions

Public detail is limited for now. It is not yet clear whether similar gains hold across other models and benchmarks, what happens over multiple refinement steps, or how much extra compute the process requires. The reported 11.5-point improvement refers to one model on one benchmark.

The permissive license does mean outside researchers can test the claims themselves. The release fits a growing view in the field that agent performance depends not only on the underlying model but also on the structure built around it, and Harness2 offers a concrete attempt to automate that part of the work.