Devtoolwatch BoardHarness² review
Harness² is a Python framework from Google Research for recursive agent harnessing, refining an agent's prompts, tools, memory and skills while model weights stay fixed.
TL;DR
Harness² is a research framework that refines the harness around a model for each new task. Harness² targets researchers running agent benchmarks such as JobBench, LAB and WorkBuddy-Bench. Harness² uses a Propose–Probe–Compose step without ground-truth feedback, and it carries lessons forward in a persistent improvement harness.
- 21.7kGitHub stars
- +2.5knew stars, week of Oct 25
- #1in Agent frameworks this week
Harness² at a glance
- Category
- Agent frameworks
- Maintained by
- Google Research (organization, @google-research on GitHub)
- License
- Apache 2.0
- Language
- Python, Shell
- Repository
- github.com/google-research/harness2
- First published
- Oct 6, 2026
What Harness² does
- Harness² refines the execution harness for the current task by proposing edits, probing their effects, and composing a refined harness from behavioral contrasts.
- Harness² updates a persistent improvement harness with adopted harnesses and evidence-supported lessons from each task to guide future refinement.
- Harness² exposes eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.
- Harness² supports parallel recursion, which explores multiple candidate harnesses, and sequential recursion, which refines through successive steps.
- Harness² runs tasks with OpenCode as the default executor, or with Codex through the --substrate cx option.
- Harness² includes setup, judging and reporting commands for the JobBench, LAB and WorkBuddy-Bench benchmarks.
- Harness² saves results under runs/, including deliverables, adopted harnesses and reflection notes. Repeating a command resumes the run.
How it works
Harness² runs as command-line tools in a Python 3.12 environment on Linux. harness2-setup installs a benchmark, and harness2 runs refinement with a --k recursion budget and a parallel or sequential mode. harness2-judge and harness2-report then evaluate the run. Model-backed runs use Vertex AI models configured in .env with Google Cloud authentication. Tasks execute through OpenCode or Codex.
How to install and start using Harness²
Requirements
- Linux, Python 3.12, and uv
- Node.js 20+ and the Google Cloud CLI for model-backed benchmarks
- A Vertex AI project and model IDs configured in .env
- Pandoc and Poppler for LAB, and Docker for WorkBuddy-Bench
First steps
- Run the local demo with harness2-demo --output runs/demo --k 2. The demo needs no model credentials.
- Copy .env.example to .env, add your Vertex AI project and model IDs, then run gcloud auth application-default login.
- Install a benchmark with harness2-setup lab, then verify it with harness2-check --bench lab.
- Add --dry-run to harness2 to preview task selection without model calls before a small run.
Who Harness² is for
- Researchers studying how to optimize agent harnesses without changing model weights
- Teams evaluating agents on professional, legal, office, web or code benchmarks
- Developers experimenting with OpenCode or Codex executors and self-refining agent setups
Limitations to know
- Harness² is not an officially supported Google product.
- Harness² model-backed runs incur API costs, so the README advises starting with a small task selection.
- Harness² executors have unrestricted network access, and Codex runs with its sandbox disabled. You must configure network restrictions yourself.
Is Harness² growing? Usage data
In the week of Oct 25–31, 2026, Harness² ranked #10 on the Devtoolwatch Board and #1 in Agent frameworks, with a momentum score of 95 out of 100.
Show the weekly numbers
| Week of | New stars |
|---|---|
| Aug 9, 2026 | – |
| Aug 16, 2026 | – |
| Aug 23, 2026 | – |
| Aug 30, 2026 | – |
| Sep 6, 2026 | – |
| Sep 13, 2026 | – |
| Sep 20, 2026 | – |
| Sep 27, 2026 | – |
| Oct 4, 2026 | 2 |
| Oct 11, 2026 | 2k |
| Oct 18, 2026 | 17.2k |
| Oct 25, 2026 | 2.5k |
On Hacker News
- Harness2 by Google Research – Recursive Agent Harnessing, 81 points and 19 comments (Oct 19, 2026)
- Harness2, 5 points and 2 comments (Oct 18, 2026)
- Google Harness2: Recursive Agent Harnessing for an Open World, 5 points and 0 comments (Oct 18, 2026)
The Board ranks tools by four signals each week: new stars, new stars for the repo's size, added downloads and download growth beyond the market. How the ranking works.
Harness² alternatives
Other AI agent frameworks on the Board in the week of Oct 25–31, 2026, by momentum:
| Rank | Tool | Stars | Downloads/wk |
|---|---|---|---|
| #2 | HyperFramesHyperFrames is an open-source framework that turns HTML, CSS, media, and seekable animations into deterministic MP4 videos, built for AI… | 76.6k | 1.5M |
| #3 | AgentScopeAgentScope is a Python agent framework with an SDK for building ReAct agents and an agent service backend with a web UI. | 35k | 51.4k |
| #4 | AG-UIAG-UI is an open, event-based protocol that standardizes how AI agents connect to user-facing frontend applications. | 16.9k | 4.6M |
| #5 | LangGraphLangGraph is a low-level orchestration framework and runtime for building, managing, and deploying long-running, stateful agents in… | 44.2k | 18.5M |
| #6 | BifrostOpen-source, high-performance AI gateway that works as your LLM gateway, MCP gateway and agent gateway. One OpenAI-compatible API for… | 9.3k | – |
| #7 | assistant-uiassistant-ui is an open-source TypeScript/React library of composable primitives and runtimes for building AI chat interfaces in your app. | 12.9k | 2.8M |
Harness² FAQ
Is Harness² free and open source?
Yes. Harness²'s source code is on GitHub (google-research/harness2) under the Apache 2.0 license, so you can use it for free.
Is Harness² still maintained?
Yes. The google-research/harness2 repository was last updated on Oct 14, 2026.
How popular is Harness²?
Harness² has 21.7k GitHub stars as of Nov 1, 2026. It gained 2.5k new stars in the week of Oct 25–31, 2026, against a 12-week average of 5.4k a week. That made it #1 among AI agent frameworks on the Devtoolwatch Board.
What are the best alternatives to Harness²?
Other AI agent frameworks on the Devtoolwatch Board in the week of Oct 25–31, 2026, by momentum: HyperFrames (#2), AgentScope (#3), AG-UI (#4), LangGraph (#5), Bifrost (#6). The alternatives table above has their numbers.
What is an agent harness in Harness²?
In Harness², a harness is the system around a model that lets it act as an agent: prompts, tools, memory, skills, orchestration, and more. Harnessing is the process of optimizing that system.
Does Harness² need ground-truth answers to improve a harness?
No. Harness² composes a refined harness from behavioral contrasts between executions on the same task, without ground-truth feedback.
What is the difference between parallel and sequential recursion in Harness²?
Harness² parallel recursion probes multiple candidate harnesses and composes refinements from their evidence. Sequential recursion carries the refined harness through successive steps. Parallel explores wider, and sequential refines deeper.
Can I try Harness² without model credentials?
Yes. The Harness² local demo runs both recursion modes with deterministic editors and real Python executions. The demo requires no model credentials.
What does the --k option control in Harness²?
In Harness², --k sets the task-level recursion budget. With --k 1, Harness² runs one Propose–Probe–Compose step. Judging and reporting must use the same --k, mode, benchmark, executor, subset and seed as the refinement run.
Sources
Harness² README badge
Maintainers can show the tool's weekly Board rank in their README:
[](https://index.devtoolwatch.com/tools/google-research-harness2/)How this page is made: Devtoolwatch writes each review from the project's README, releases, package metadata and website, with an AI model, then checks every claim against the source it cites and drops what the sources don't support. The numbers come from Devtoolwatch's daily GitHub snapshots, GitHub's stargazer history, and the npm and PyPI download APIs. Spotted an error? Email [email protected].