Google open-sources Harness2 for test-time agent harness rewriting

On October 18, 2026, Google Research open-sourced Harness2, a framework released under the Apache-2.0 license. According to MarkTechPost, Harness2 can rewrite an agent's harness at test time. A single refinement step raised GLM-5 by 11.5 points on WorkBuddy-Bench, with no model weight updates involved.

The following day, October 19, a user on the LocalLLaMA subreddit offered a fuller description of what the release does. The post describes a harness as everything surrounding the model, including the system prompt, tools, memory, skills, and subagents. Normally a person designs this setup once, and the agent then reuses that fixed configuration for every new task. With Harness2, the agent refines the harness itself for each task: it proposes changes, runs several candidate versions, and selects a better one by comparing how they perform, without needing correct answers or grades.

The two accounts agree on the release date, the open-source publisher, and the license, but they emphasize different aspects. MarkTechPost supplies a concrete figure, the 11.5-point gain from one refinement step, while the Reddit post focuses on the mechanism of per-task self-improvement and comparison across candidates, without quantitative results. Public information so far remains at the stage of the open-source release and community discussion.