Google Open-Sources Harness²: A Recursive Agent-Harnessing Framework
Google Cloud AI Research has open-sourced Harness², a recursive agent-harnessing framework that improves the code and files surrounding a frozen model. For developers, the pitch is straightforward: your agent may get better without swapping in a better model.
TL;DR
- What it is: A framework, not a model. Zero parameters are trained; instead, eight editable harness components are optimized. Current version: 0.1.0.
- Runs on: Linux with Python 3.12 or 3.13 and
uv. Model calls are routed through Vertex AI. GPU requirements: not disclosed. - Performance: A single refinement step improved all four model-and-harness setups across all three benchmarks.
- Best result: GLM-5 running under OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench.
Why It Matters
In 2026, agent performance is increasingly shaped by the scaffolding around a model—prompts, tool definitions, orchestration logic, and supporting files—rather than by the model weights alone. Harness² leans into that reality: it treats the harness itself as a searchable, improvable artifact while keeping the underlying model frozen. This makes it a practical option for teams that cannot fine-tune or retrain large models but still need measurable gains in agentic tasks.
Key Concepts
- Frozen model, editable harness: The framework iteratively refines eight harness components instead of touching model parameters.
- Recursive refinement: Each refinement step feeds improvements back into the harness, compounding gains across benchmarks.
- Model-agnostic: Tested across multiple model-and-harness pairings, with consistent improvements reported.
via MarkTechPost