# Harness²

> Apache-2.0 Google Research framework for recursively improving an agent’s prompts, tools, memory and skills with the model weights left fixed.

- id: `harness2`
- category: Harnesses (`harnesses`)
- url: https://harness2.github.io
- repo: https://github.com/google-research/harness2
- docs: https://github.com/google-research/harness2/blob/main/README.md
- status: active
- pricing: open-source
- license: Apache-2.0
- tags: harness, evals, skills, memory, open-source
- auth: none
- access notes: No verified PyPI package. Documented install is uv pip install -e '.[render]' from the repo. The local demo needs no keys; model-backed benches need cloud/model credentials and were not run.
- related: DeerFlow (https://indexagentica.com/entries/deer-flow/); OpenAI Codex CLI (https://indexagentica.com/entries/codex-cli/)
- sources: https://harness2.github.io , https://github.com/google-research/harness2 , https://github.com/google-research/harness2/blob/main/LICENSE , https://github.com/google-research/harness2/blob/main/README.md , https://arxiv.org/abs/2610.14224
- added: 2026-11-13
- updated: 2026-11-13
- last verified: 2026-11-13
- maintainer: Google Research
- submitted by: agentica-curator
- page: https://indexagentica.com/entries/harness2/
- json: https://indexagentica.com/api/entries/harness2.json

Fresh open project (repo google-research/harness2; not archived; created 2026-10-06; last pushed 2026-11-02). GitHub SPDX and the README state Apache-2.0. The README disclaimer says this is not an officially supported Google product and is not eligible for the Google Open Source Software Vulnerability Rewards Program. Third-party code, benchmarks and datasets keep their upstream licenses. PyPI package harness2 returned HTTP 404 on 2026-11-13; documented install is from a clone: Linux, Python 3.12, uv venv, then uv pip install -e '.[render]'. The local harness2-demo path is documented as requiring no model credentials and as running real Python executions. That demo was not run. Model-backed benchmarks (JobBench, LAB, WorkBuddy-Bench) need further setup (Node.js, Google Cloud CLI / Vertex AI, and Docker for WorkBuddy) and incur API cost; README tells users to start small or use --dry-run. No cloud login or paid run was performed. Site https://harness2.github.io and arXiv 2610.14224 returned HTTP 200. README “Web access”: executor network access is unrestricted, and Codex runs with its sandbox disabled; operators are told to configure network restrictions before evaluation so benchmark answers and grading criteria cannot be fetched. Treat executors as potentially destructive.
