Google Cloud AI Research has open-sourced Harness², a recursive agent harnessing framework designed to improve AI agent performance by optimizing system prompts, tool interfaces, and execution logic around a frozen model. The framework operates on unseen open-world tasks without relying on external verifiers or ground-truth labels.
Harness² structures the operational environment into eight editable components, expanding on prior agent engineering layouts. It implements parallel and sequential recursive search modes to automatically refine agent harnesses, enabling weaker executor models like GLM-5 to outperform stronger base models on professional benchmarks such as WorkBuddy-Bench and Harvey’s LAB.
By decoupling agent performance gains from underlying model parameter updates, Google’s framework allows engineering teams to optimize execution flows locally. The open-source code is hosted on GitHub, targeting Linux environments with Python 3.12 and requiring Node.js and Vertex AI infrastructure for model-backed deployments.
Why it matters
Developers can improve agent performance and step efficiency on complex tasks without re-training or fine-tuning foundation models.
Open-source harness engineering provides startups with a structured, label-free optimization framework for production agent workflows.
Source: marktechpost.com



