Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

Google Research has released Harness2, an open-source framework licensed under Apache-2.0 that takes a recursive approach to agent harnessing. Rather than modifying model weights, Harness2 rewrites the harness — the scaffolding that wraps an agent's reasoning and tool-use loop — at test time, allowing the system to adapt its own orchestration logic on the fly.

The framework's core idea is that the harness itself can be treated as a mutable artifact. By recursively refining how an agent plans, calls tools, and evaluates outputs, Harness2 can improve performance without any fine-tuning or weight updates. In reported results, a single refinement step lifted GLM-5's score by 11.5 points on the WorkBuddy-Bench benchmark.

This matters because agent performance is often bottlenecked not by the underlying model's raw capability but by how effectively that capability is orchestrated. A framework that can self-improve its own orchestration layer at inference time could reduce the need for expensive retraining cycles and make agent systems more adaptable to new tasks.

The Apache-2.0 license makes Harness2 accessible to the broader research and developer community. Whether the reported gains generalize across different base models and benchmarks will require further independent validation, but the approach of treating the harness as a first-class, optimizable component is a notable direction in agent research.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.