Harness2 (Harness²) is Google Research's open-source framework for recursive agent harnessing.
Follow Harness2 to get an email when its listing changes, such as new pricing or features. One digest a day at most, and you can unsubscribe anytime.
Model weights stay fixed while it refines the harness around coding agents such as OpenCode and Codex for each task, and it accumulates lessons in a persistent improvement harness for tasks it has not seen before.
- 01
Task-level Propose-Probe-Compose refinement without ground-truth feedback
- 02
Persistent improvement harness that keeps adopted harnesses and lessons
- 03
Eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins
- 04
Parallel and sequential recursion modes with a configurable budget
- 05
OpenCode and Codex executors
- 06
JobBench, LAB and WorkBuddy-Bench setup, judging and reporting CLIs
Open Source
Free and open source under Apache-2.0. The local demo needs no model credentials; benchmark runs call Vertex AI models and incur your own API costs.
Best for
Researchers and agent builders who want to automatically tune prompts, skills, tools and guardrails around coding agents and measure the gains on benchmarks.
Not ideal for
Developers looking for a ready-to-use coding assistant; it is a Linux and Python 3.12 research framework that needs Vertex AI setup for model-backed runs.
A notable research release that treats the agent harness, not the model, as the thing to optimize, with practical OpenCode and Codex integration for experimentation.
Open-sourced by Google Research on October 8, 2026 with an accompanying arXiv paper; the repository was updated on October 14, 2026 and reached the Hacker News front page on October 19, 2026.
As of
Is Harness2 free?+
Yes - Harness2 is Open Source. Free and open source under Apache-2.0. The local demo needs no model credentials; benchmark runs call Vertex AI models and incur your own API costs.
Is Harness2 open source?+
Yes - Harness2 is open source. Free and open source under Apache-2.0. The local demo needs no model credentials; benchmark runs call Vertex AI models and incur your own API costs.
Who is Harness2 best for?+
Researchers and agent builders who want to automatically tune prompts, skills, tools and guardrails around coding agents and measure the gains on benchmarks.
Who is Harness2 not ideal for?+
Developers looking for a ready-to-use coding assistant; it is a Linux and Python 3.12 research framework that needs Vertex AI setup for model-backed runs.
What are the best Harness2 alternatives?+
The closest Harness2 alternatives on ai.dosa.dev are Pieces for Developers, agenttrace, Vibe Kanban - all listed under Developer Productivity & Workflow.
Who makes Harness2?+
Harness2 is developed by Google Research. It is listed in the Developer Productivity & Workflow category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.