Google Research Open-Sources Harness2 Framework for Recursive Agent Optimization
Google released Harness2, an Apache-2.0 licensed framework that improves agent performance by refining external harness components without updating model weights.
Key points
- Harness2 uses recursive refinement to improve agent performance without changing underlying model parameters.
- Testing showed GLM-5 performance on WorkBuddy-Bench increased by 11.5 points with one refinement step.
- The framework operates without ground-truth labels or verifiers, targeting open-world task scenarios.
Google Cloud AI Research has open-sourced Harness2, a framework designed to optimize the operational environment of artificial intelligence agents. Released under the Apache-2.0 license, the tool allows developers to enhance agent performance by modifying the harness components that surround a frozen model, rather than retraining the model itself.
The framework targets the "open world" setting where tasks are unseen and no verifier confirms answer correctness. According to MarkTechPost, Harness2 improves the code and files around a model at test time. One refinement step lifted GLM-5 performance by 11.5 points on the WorkBuddy-Bench benchmark without any weight updates.
Technical Architecture and Performance
Harness2 is a framework, not a model, containing zero trained parameters. It consists of eight editable harness components, each mounted as a separate file. These components cover context construction, tool interfaces, delegation, and behavior within the tool loop. The layout extends a previous seven-file split from earlier Agentic Harness Engineering work by adding one component.
The system runs on Linux with Python 3.12 or 3.13 and uv. Model runs call models through Vertex AI, though specific GPU requirements were not disclosed. The framework exposes eight editable components that allow for plug-and-play refinement during deployment. A harness includes prompts, tools, memory, skills, and orchestration logic.
Performance varied across benchmarks. The best result saw GLM-5 under OpenCode rise from 64.2 to 75.7 on WorkBuddy-Bench. The worst result showed GLM-5 under Codex gaining only 1.2 points on JobBench, moving from 23.6 to 24.8. All four model and harness setups improved on all three benchmarks with just one refinement step.
Recursive Refinement Process
Harness2 employs two linked levels of recursion to optimize agent behavior. Task-level recursion uses a contrastive Propose, Probe, Compose step. The agent first runs the task with the base harness, then an editor proposes two candidates. One candidate targets general failures like planning, while the other targets domain-specific knowledge.
The composer compares trajectories against the reference and keeps only edits backed by observed behavior. Harnessing-level recursion occurs after each task. A reflector stores the adopted harness and an EXPERIENCE.md file of lessons. This library guides future proposals, ensuring each new task starts from the base harness but benefits from accumulated experience.
What should you do?
Developers interested in optimizing agent performance without retraining models should review the Harness2 documentation. Teams using Vertex AI can test the framework with Python 3.12 or 3.13 environments. Organizations should evaluate whether the eight editable components align with their specific agent orchestration needs.
What to do and how to stay safe: Google
- Review the Apache-2.0 licensed codebase to understand the eight editable harness components.
- Test the framework in a controlled environment using Python 3.12 or 3.13 on Linux.
- Monitor performance changes on specific benchmarks like WorkBuddy-Bench or JobBench.
- Ensure Vertex AI integration is configured correctly for model runs.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Harness2 require retraining the underlying AI model?
No, Harness2 works with frozen models and never changes model weights, focusing instead on optimizing the surrounding harness components.
What benchmarks showed the most significant improvement?
GLM-5 under OpenCode saw the largest gain, rising 11.5 points on WorkBuddy-Bench, while improvements on JobBench were smaller.
What are the system requirements for running Harness2?
The framework runs on Linux with Python 3.12 or 3.13 and uv, and calls models through Vertex AI.
