Google Research has published Harness2, an Apache 2.0 framework that rewrites the instructions and tools around an AI agent for each task it is given, then saves what worked for the next one. The model itself is never retrained. Its weights stay fixed throughout.
That limit matters because the project’s own name for the idea is “recursive agent harnessing”, and a harness is only the scaffolding that surrounds a model: plugins, scripts, skills, subagents, memory, guardrails, tool guidance, and the system prompt. Harness2 edits those eight components. It does not make a model smarter, and nothing in the repository’s README claims otherwise. What it automates is the trial-and-error tuning most teams building agents currently do by hand.
The loop works at two levels. On a single task, the system proposes edits to the harness, runs them against that same task to see what changes, and composes a refined version from the differences in behavior between runs. The README says this happens without ground-truth feedback, so no answer key is needed while the agent works. Across tasks, a separate persistent harness collects the versions that were adopted along with lessons the README describes as “evidence-supported.”
What those lessons look like on disk is thinly documented. The README says a run saves reflection notes, adopted harnesses, deliverables, and task summaries, but it does not describe how a later task retrieves a lesson or how the system decides one is supported. That is the part to read in the paper (arXiv 2610.14224, by Ruiyao Xu and colleagues, including co-author Tomas Pfister) before trusting it.
The numbers come from three professional-work benchmarks: JobBench, LAB (legal work), and WorkBuddy-Bench (code, office, and web tasks). With one refinement step, the README reports that GLM-5 and Gemini 3.5 Flash both beat their unmodified starting harnesses, by up to 4.5 points on JobBench, 11.5 points on WorkBuddy-Bench, and, scored by partial pass rate, 11.8 percentage points on LAB. Those are best-case ceilings, and the README does not say which model produced which. It adds that giving the system more refinement attempts per task helps further, through either parallel candidates or successive passes.
The open question for builders is whether the saved lessons generalize or just fit the tasks already seen. The README frames the target as an “open world” where tasks were not known when the harness was designed, which is the right test, but the page itself shows only benchmark gains, not transfer to unrelated work.
The README does flag one hazard. Executors can reach the network without restriction, Codex’s sandbox is turned off, and public benchmark rubrics may be readable online, so users must restrict the network before evaluating or an agent could find the answers. Setup is also heavier than a pip install: Linux, Python 3.12, a Google Cloud login, Docker for one benchmark, and paid API calls for model-backed runs. A free local demo runs without credentials. Google also states it is not an officially supported product.
For teams hand-tuning agent prompts, the practical test is cheap: run it on one benchmark task, read the adopted harness it produces, and judge whether the edits are ones you would have made yourself.
Based on the google-research/harness2 repository README on GitHub, read on 15 October 2026.
