Harness²

Harness²

Harness² is an Apache-2.0 research framework from Google Cloud AI Research that automatically refines an LLM agent's harness (prompts, tools, memory, skills and more) for each task and carries lessons forward to future tasks, without changing model weights.

0.0 (0)
2 Impressions
Visit Website
Founder
Google Research
Active
PlanOpen source - $0 - Apache 2.0; model-backed runs incur your own API costs
BusinessB2B

About Harness²

What is Harness²?

An Apache-2.0 framework from Google Cloud AI Research that automatically refines an LLM agent's harness for each task and learns how to refine better over time, with model weights fixed.

How it works

  • Task-level recursion: Propose-Probe-Compose edits to eight harness components, parallel or sequential
  • Harnessing-level recursion: a persistent improvement library and EXPERIENCE.md guide future tasks
  • OpenCode and Codex executors; JobBench, LAB and WorkBuddy-Bench setups

Results (reported by the authors)

Up to +11.5 points on WorkBuddy-Bench, +4.5 on JobBench and +11.8 pp on LAB partial pass rate from one step with GLM-5 and Gemini 3.5 Flash.

Getting started

Linux, Python 3.12 and uv; a credential-free local demo is included. Free and open source; not an officially supported Google product.

Value & Audience

Value Proposition

Gives agent builders an automated, evidence-driven way to improve prompts, tools, memory and skills around a fixed model instead of hand-tuning the harness for every new kind of task.

Problem Solved

Agent harnesses are usually hand-designed and break down on tasks their designers did not anticipate, and most automatic optimizers need a verifier or ground truth; Harness² refines the harness from behavioral contrasts between executions.

Share this product

Product founder?

Get Badge

Tech Stack

Social & Links

Loading ratings...

Loading comments...

Gallery

Harness² screenshot 1

Feedback & Roadmap

Loading feedback…

Harness² FAQ

Common questions about Harness²'s features, pricing, and use cases

What does Harness² do?

Open-source framework that recursively tunes agent harnesses Gives agent builders an automated, evidence-driven way to improve prompts, tools, memory and skills around a fixed model instead of hand-tuning the harness for every new kind of task.

Is Harness² free to use?

Harness² is open source and free to use, in the AI Agents & Automation category.

What problem does Harness² solve?

Agent harnesses are usually hand-designed and break down on tasks their designers did not anticipate, and most automatic optimizers need a verifier or ground truth; Harness² refines the harness from behavioral contrasts between executions.

Who is Harness² built for?

Harness² is built for Data Scientists & ML Engineers, Researchers, Developers, Open Source Maintainers.

What tech stack does Harness² use?

Harness² is built with Python, uv, OpenCode, Codex, Vertex AI.

What category is Harness² listed under?

Harness² is listed in the AI Agents & Automation category on TechLogHub, alongside curated alternatives with community ratings and pricing comparisons.

Harness² appears in

Newsletter

One email a week

TechLogHub tracks 606 tools and 535 open-source projects. Get the week's new launches, release roundups and open-source picks — nothing else.

Get the TechLogHub digest

One email a week: new and trending developer tools, fresh comparisons, and what shipped. Unsubscribe in one click.