Google Cloud AI Research has open-sourced Harness², a recursive agent harnessing framework that improves the code and files around a frozen model. The framework aims to enhance agent performance without requiring changes to the model itself.
Key points
- Google open-sources Harness2, a recursive agent harnessing framework.
- Harness2 improves AI agents by refining surrounding components.
- Model weights remain unchanged during the harnessing process.
- The framework operates without labels or test cases.
- Eight editable components make up the Harness2 framework.
Harness² is a plug-and-play framework from Google Research that refines an agent’s harness during deployment. A harness encompasses elements such as prompts, tools, memory, skills, and orchestration. Harnessing is the process of optimizing this system, with model weights remaining unchanged.
The framework is designed for open-world settings where tasks are unseen and no verifier confirms answer correctness. Unlike most self-improvement methods that assume labels or test cases, Harness² operates without them.
By the numbers
- GLM-5 improvement on WorkBuddy-Bench
- +11.5 points
- GLM-5 gain on JobBench
- 1.2 points
Harness² exposes eight editable components, each mounted as a file: System prompt, Guardrails, Memory, Tool guidance, Subagents, Skills, Scripts, and Plugins (middleware). These components manage context construction, tool interface, delegation, and behavior within the tool loop.
The framework utilizes two linked levels of recursion. Task-level recursion employs a contrastive Propose, Probe, Compose step. Initially, the agent executes the task with the base harness, followed by an editor proposing two candidates. The invariant candidate addresses general agent failures like planning or tool use.
Sources and references
Original report
Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Paperarxiv.org/abs/2610.14224 (opens in a new tab)
- Codegithub.com/google-research/harness2 (opens in a new tab)
- Codegithub.com/google-research/harness2/blob/main/pypr… (opens in a new tab)
- Paperarxiv.org/abs/2604.25850 (opens in a new tab)
- Paperarxiv.org/abs/2607.20911 (opens in a new tab)
- Paperarxiv.org/abs/2605.26329 (opens in a new tab)
Ref 00SVVKF4
Cite this
APA
Interestana AI Editorial. (2026, October 18). Google Open-Sources Harness2 Recursive Agent Framework. Interestana. https://interestana.com/articles/google-open-sources-harness2-a-recursive-agent-harnessing-framework-00svvkf4
MLA
“Google Open-Sources Harness2 Recursive Agent Framework.” Interestana, 18 Oct. 2026, interestana.com/articles/google-open-sources-harness2-a-recursive-agent-harnessing-framework-00svvkf4.
Chicago
Interestana AI Editorial. “Google Open-Sources Harness2 Recursive Agent Framework.” Interestana, October 18, 2026. https://interestana.com/articles/google-open-sources-harness2-a-recursive-agent-harnessing-framework-00svvkf4.
BibTeX
@misc{interestana_00svvkf4,
author = {{Interestana AI Editorial}},
title = {{Google Open-Sources Harness2 Recursive Agent Framework}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/google-open-sources-harness2-a-recursive-agent-harnessing-framework-00svvkf4},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/18/google-harness2-recursive-agent-harnessing-framework/}
}Credit the source
Original reporting: MarkTechPost, “Google Open-Sources Harness2, a Recursive Agent Harnessing Framework”, https://www.marktechpost.com/2026/10/18/google-harness2-recursive-agent-harnessing-framework/(opens in a new tab)
Follow this story
Get Computing & Chips in your inbox — new stories in 40 words or fewer, once a week.
Subscribe