Harness²

What makes it different

Harness² is an open-source, plug-and-play framework from Google Cloud AI Research (with Northwestern University) for recursive agent harnessing in an open world. The problem it attacks: LLM agents live or die by their execution harness — the prompts, tools, memory, skills, guardrails and orchestration wrapped around the model — and today that harness is hand-built. Harness² automatically evolves it, without retraining model weights and without ground-truth verifiers.

Two levels of recursion drive it. At the task level, a contrastive Propose–Probe–Compose step refines how the agent executes a task: propose candidate approaches, probe them, compose the best into a working harness. At the harnessing level, reusable experience accumulates across tasks, so the system gets better at building harnesses the more it runs. The model stays frozen; the scaffolding around it is what learns.

The numbers come from the paper "Harness²: Recursive Agent Harnessing for an Open World" (arXiv:2610.14224): a single refinement step improved GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 points on LAB. It is released as v0.1.0 under Apache 2.0, with 8 editable harness components and support for OpenCode and Codex executors, running on Vertex AI. Independent deep-dives add the honest caveats: multi-candidate probing multiplies test-time token costs, and the main experiments rely on a frontier editor model.

Harness² is a research release — the paper lists 9 authors with Ruiyao Xu as first author — and the code is open at github.com/google-research/harness2. For teams building agents on frozen models, it is one of the more concrete attempts to make the harness itself a first-class, improvable artifact.