New tool2026-10-19via Product HuntAI Agents That Do Tasks for You

Harness2

Name: Harness2
Author: Google Research

Harness2 is an open-source research framework from Google Research that improves the setup around an AI agent (prompts, tools, memory, skills) task by task, without changing model weights, and carries lessons forward to later tasks.

Visit Harness2 ↗Repository ↗

Free (open source)Apache-2.0Python (3.12, uv; Node.js 20+ and Google Cloud CLI for model-backed benchmarks)★ 13,815 (as of 2026-10-20)by Google Research

Why it matters

For AI researchers and teams building agents who want to test automatic harness tuning. It targets tasks the harness designers never saw, using a propose, probe and compose loop that doesn't need ground-truth answers.

Highlights

Task-level recursion proposes edits to the harness, probes their effect on the same task, and composes a refined harness from the differences.
Harnessing-level recursion keeps a persistent improvement harness that stores adopted harnesses and lessons for future tasks.
Edits eight components: the agent's main instructions, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.
Supports parallel and sequential recursion, OpenCode and Codex executors, and the JobBench, LAB and WorkBuddy-Bench benchmarks.
The authors report gains of up to 11.5 points on WorkBuddy-Bench and 4.5 points on JobBench for GLM-5 and Gemini 3.5 Flash after a single step.unverified

Taken from the project's own README, docs or launch page. Claims about speed or accuracy are the maker's unless stated otherwise.

Try it if: You build or study AI agents and want to see whether automatic harness tuning beats your hand-written prompts and tools.

Worth knowing: Research code, labeled not an officially supported Google product. Executors run with unrestricted network access and Codex runs with its sandbox disabled, so use an isolated environment. Benchmarks and third-party code keep their own licenses.

Quick facts

Pricing	Free (model-backed runs incur your own API costs)
Free or paid	Free (open source)
License	Apache-2.0
Language	Python (3.12, uv; Node.js 20+ and Google Cloud CLI for model-backed benchmarks)
Platform	Linux; Vertex AI for model-backed runs
GitHub stars	★ 13,815 (as of 2026-10-20)
Maker	Google Research

Related ranking

See how established tools compare in Best AI Agents That Do Tasks for You. New releases aren't scored until we review them against the rubric.

What we couldn't verify

highlights[4] (benchmark gains are the authors' own results)

Checked 2026-10-20 from the tool's own pages. Prices and star counts change, so confirm on the official site. Listing here is not a ranking or an endorsement.