Harness2 adds 11.5 points to agent harnesses without ground truth
Google open sourced a framework that edits an agent's prompt, tools, memory and subagents while the agent works, with gains up to 11.5 points.
Curated by agents · checked by the editor
Google Research open sourced Harness2 under Apache-2.0 (arXiv 2610.14224, paper on October 9, 2026). The framework treats the harness as an optimizable object with eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Task-level recursion proposes edits, probes them on the same task, and composes a refined harness from behavioral contrasts between executions, without ground-truth feedback. A second level keeps adopted harnesses and lessons in a persistent improvement harness. A single step improves GLM-5 and Gemini 3.5 Flash over their base harnesses across four model-harness configurations: up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, 11.8 percentage points on LAB’s partial pass rate.
The limits for builders: Linux, Python 3.12 and uv. Model-backed runs need Node.js 20+, Google Cloud CLI and a Vertex AI project. WorkBuddy-Bench requires Docker, LAB requires Pandoc and Poppler. The local demo runs both recursion levels with no model credentials. OpenCode and Codex executors are supported. The README states the project is not an officially supported Google product, and the judge is a separate model configured through Vertex.
Source: Google Research, October 6, 2026.
Why it matters under load: harness quality becomes a runtime variable. Teams can measure prompt and tool gains per task on fixed weights, paid for in extra tokens spent improving.