Harness²
Harness² (harness2) is a Google Research plug-and-play framework for recursive agent harnessing in an open world. It keeps model weights fixed while optimizing the harness around a model across two levels of recursion — task-level and harnessing-level — through eight editable harness components. Released under the Apache 2.0 license.
Key facts
What is Harness²?
Improves agent performance by recursively optimizing the harness around a frozen model rather than fine-tuning model weights.
Who is Harness² best for?
AI researchers, agent developers, ML engineers working with LLM agents on Linux with Python 3.12
What are its main limitations?
- Not an officially supported Google product and not eligible for the Google Open Source Software Vulnerability Rewards Program
- Benchmark figures are maximums across three benchmarks and four model–harness configurations, not typical or absolute scores
- Documentation is silent on some details such as GPU requirements, guardrail implementations and the skill file format
- Benchmark harnesses have external dependencies (Pandoc and Poppler for LAB, Docker for WorkBuddy-Bench)
Limitations are based on product materials or editorial synthesis; confirm on the official site.
Key features
- ✓Two levels of recursion: task-level recursion refines the execution harness via a contrastive Propose–Probe–Compose step, and harnessing-level recursion writes adopted harnesses plus evidence-supported lessons into a persistent improvement harness for future tasks
- ✓Eight editable harness components grouped by runtime role: system prompt, guardrails, memory, skills, tool guidance, subagents, scripts and plugins
- ✓Keeps model weights fixed — improves the harness rather than performing SFT or weight fine-tuning
- ✓Credential-free local demo (`harness2-demo --output runs/demo --k 2`) that runs both recursion modes with deterministic editors and real Python executions on CPU
- ✓Documented setup for model-backed runs using Vertex Gemini executors and judges, with Vertex Claude or Gemini as the improver
- ✓OpenCode as the default executor, with Codex available via `--substrate cx`
- ✓Benchmark harnesses for professional work (JobBench), legal work (LAB) and office/web/code tasks (WorkBuddy-Bench)
Use cases
- →Optimizing an agent harness for a specific task using task-level recursion
- →Persisting harness improvements and lessons across tasks via the improvement harness
- →Reproducing professional benchmark results on JobBench, LAB and WorkBuddy-Bench
- →Verifying an installation locally with the credential-free CPU demo before connecting Vertex AI
- →Comparing harness-level optimization with frameworks such as LangChain and DSPy
Pros
- +Apache 2.0 licensed Google Research project with a public repository and arXiv paper
- +Model-agnostic in principle — model weights stay fixed and only the harness is changed
- +Credential-free demo allows install verification without model access
- +No documented GPU requirement for the local demo
- +Reported gains from a single Harness² step: up to +4.5 points on JobBench, +11.8 percentage points partial pass rate on LAB, and +11.5 points on WorkBuddy-Bench
Cons
- −Not an officially supported Google product and not eligible for the Google Open Source Software Vulnerability Rewards Program
- −Benchmark figures are maximums across three benchmarks and four model–harness configurations, not typical or absolute scores
- −Documentation is silent on some details such as GPU requirements, guardrail implementations and the skill file format
- −Benchmark harnesses have external dependencies (Pandoc and Poppler for LAB, Docker for WorkBuddy-Bench)
- −README targets Linux and Python 3.12 with uv
Positioning
- Core value: Improves agent performance by recursively optimizing the harness around a frozen model rather than fine-tuning model weights.
- Ideal for: AI researchers, agent developers, ML engineers working with LLM agents on Linux with Python 3.12
- Product type: Open-source Python repository (Apache 2.0) with CLI; accompanied by an unofficial community field-guide website
Frequently asked questions
What is Harness²?
Harness² (harness2) is a Google Research plug-and-play framework for recursive agent harnessing in an open world. It connects two levels of recursion while keeping model weights fixed, optimizing the harness — prompts, tools, memory, skills, orchestration and more — rather than the model itself.
How much does Harness² cost?
No pricing plans are published; the project is released under the Apache 2.0 license according to the repository and this guide.
Do I need a GPU to run Harness²?
The official docs do not list a GPU requirement. The local demo runs on CPU with deterministic editors, and model-backed runs call the configured Vertex AI models.
Does Harness² fine-tune the model?
No. The README states that Harness² keeps model weights fixed and improves the harness instead, so it is not SFT or weight fine-tuning.
Which models and executors are documented?
Reported results cover GLM-5 and Gemini 3.5 Flash. Setup docs describe Vertex Gemini executors and judges, with Vertex Claude or Gemini as the improver; OpenCode is the default executor and Codex is used with --substrate cx.
Is Harness² an officially supported Google product?
No. The README states it is not an officially supported Google product and is not eligible for the Google Open Source Software Vulnerability Rewards Program.
You may also like
Curated suggestions based on similar categories.
Asimov 1
Asimov 1 is a 1.2 m, 35 kg humanoid robot built for whole-body control and locomotion research. It is sold as developer-friendly, open hardware that supports sim-to-real policy transfer, raw data collection, and live teleoperation via joystick, code, or an agent. Orders are started with a $499 deposit, with full pricing handled through sales or invoice request.
Exa 3.0
Exa 3.0 is the latest release of Exa's web search API for AI applications, featuring a larger embedding model, generally available Dynamic Highlights, and a substantially expanded index covering ecommerce, local businesses, news, sports, weather, flights, events, transit, and podcasts. The release introduces Exa Lite, a lower-cost search type priced at $4 per 1,000 requests, alongside Exa Auto at $7 per 1,000 requests.
RunInfra
RunInfra is a RightNow product offering hosted APIs for open models, built for AI agents. It uses session-aware routing to keep repeat calls on warm replicas, reporting high cache hit rates and lower cached-input pricing, with one API key compatible with the OpenAI and Anthropic SDKs.
Beam.win
Beam is an AI platform built for human services teams, offering AI-assisted case notes (Notes), live two-way translation (Interpret), and voice agents that handle inbound calls (Talk). It is used by frontline government, non-profit and welfare organisations to reduce administrative workload.
Step 5 Preview
Step 5 Preview is StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 600B total parameters (27B active per token), a 1M-token context window, and vision input. It targets software engineering, long-horizon agentic tasks, professional knowledge work, and finance, and is available through StepFun's products and API, with open weights planned for October 15.
Nemotron 3.5 Super VL
Nemotron 3.5 Super VL is a member of NVIDIA's Nemotron 3 family of open models with open weights, training data, and recipes. It combines visual understanding, reasoning, and tool use in a single model and is described as delivering highest in-class efficiency and accuracy for specialized domains. It can be deployed from local infrastructure to the cloud and is accessible via OpenRouter, NVIDIA NIM API, and Hugging Face.