Harness2

google-research/harness2Reviewed by AI

What it does

Harness2 is research code from a Google team for improving the system around an AI model: the prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. For each task it proposes edits, tries them out, compares the runs and composes a refined setup, keeping the model itself unchanged, then carries the lessons to the next task. It comes with three professional benchmarks (JobBench, LAB, WorkBuddy-Bench) and runs agents through OpenCode or Codex. It is for researchers: Linux, Python 3.12, uv, Node, Docker and a Vertex AI account are needed, and real runs cost model money.

How easy it is to install

D · terminal, several tools or a specific graphics card · Linux

Linux is the only system it targets ("Use Linux, Python 3.12, and uv", and the package is marked POSIX-only), and it is a D there: you create a virtual environment with uv, and the benchmarks additionally need Node.js 20+, the Google Cloud CLI, Docker for WorkBuddy-Bench and Pandoc and Poppler for LAB, all driven from the terminal. Windows and Mac are not covered.

Coming soon: install in 1 click with the RepoPicks app.

Numbers on GitHub

Stars: 21,675on Oct 31, 2026
Downloads: —all releases
Growth: +2,002stars in 6 days

Stars over time

Facts

Runs on: Linux
You need to install: Python, Node.js, Docker, Other tools
To build from source: Python
Disk space: 100 MB–300 MB · estimate
Accounts or keys: Google Cloud Vertex AI (model access)
Cost: Free, API costs extra · The code is free (Apache-2.0). The README warns that "Model-backed runs incur API costs" through your own Vertex AI account; only the built-in demo runs without model credentials.
License: Apache-2.0 (Permissive)
Can replace: LangSmith
Last update: Oct 14, 2026

Risk signals (6)

AttentionVery new repository with many stars
20552 stars in 21 days
automatic check
Attentionagents run with no sandbox and open network
README, Web access: "Executor network access is unrestricted, and Codex runs with its sandbox disabled." The project asks you to configure network restrictions yourself before evaluating, so by default the AI agents it runs on your machine have full internet access and no sandbox.
AI review
Infonot an official google product
README Disclaimer: "This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program." The repository sits in the google-research organisation, which can give a different impression.
AI review
Inforesults come from the authors paper
The gains quoted in the README ("up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate") come from the authors' own arXiv paper 2610.14224, measured with GLM-5 and Gemini 3.5 Flash.
AI review
Infomodel bills can add up
README, Run a benchmark: "Model-backed runs incur API costs; start with a small task selection", and the recursion budget --k multiplies the number of agent runs per task.
AI review
Infoearly code with no releases
Version 0.1.0 in pyproject.toml, no GitHub releases, created 2026-10-06 with the last commit on 2026-10-14, while the checks already flag 20552 stars in 21 days.
AI review

Evidence is shown in its original language.

What Radar could not check

Whether the reported gains reproduce with other models or on your own tasks.
What a full benchmark sweep really costs in model usage.
How much disk space the three benchmarks and their Docker images take.
Whether the benchmark data and third-party code under third_party/ and dependencies/ can be used for your purpose; they keep their own licences.
What the agents do with unrestricted internet access during a run.