Google Cloud AI Research open-sources Harness² recursive agent framework
Google Cloud AI Research released Harness², an open‑source recursive agent harnessing framework version 0.1.0. The framework adds eight editable harness components—system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins—while keeping model weights frozen. It runs on Linux with Python 3.12 or 3.13 and uses Vertex AI to call models; GPU requirements are not…
Key points
- Harness² provides eight editable harness components and refines agents without changing model weights.
- One refinement step raised GLM‑5 (OpenCode) from 64.2 to 75.7 on WorkBuddy‑Bench (+11.5 points).
- Runs on Linux with Python 3.12/3.13, uses Vertex AI for model calls; GPU needs not disclosed.
The research team evaluated Harness² on three professional benchmarks (WorkBuddy‑Bench, JobBench, Harvey’s LAB) using Gemini 3.5 Flash, GLM‑5 under OpenCode and Codex, with Claude Opus 4.8 as the default editor. A single refinement step lifted GLM‑5 (OpenCode) from 64.2 to 75.7 on WorkBuddy‑Bench (+11.5 points), while the same model under Codex improved only 1.2 points (23.6 to 24.8). Parallel recursion with k=5 raised Gemini (OpenCode) to 80.8 on WorkBuddy‑Bench. The framework supports both parallel and sequential recursion modes, storing edits in an EXPERIENCE.md library for future tasks.
The story so far
3 episodes →- Google Cloud AI Research open-sources Harness² recursive agent frameworkthis story
Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
MarkTechPost · 18 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Anthropic is reportedly consulting law firms on negotiation tactics for its AI agents · 2 src
- Six AI agents spam MP office and leak data while chasing £350 train refund · 1 src
- Things to Do API launches pay‑as‑you‑go pricing for event queries · 1 src
- Voxel51 launches 3.0 platform for physical AI data workflows · 2 src
- Claude Code team shares how agents review code with minimal human input · 3 src
Comments
via GitHub Discussions