Harness²

Harness² is an agent research framework released by Google Research, with an accompanying paper. The README calls everything around a model that lets it act as an agent (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins) a harness. Harness² studies how to improve that harness automatically when an agent meets tasks it wasn't designed for, while the model weights stay fixed. It works at two levels. On the current task, it proposes several edits, probes each one, and composes a refined harness from the differences between runs. After each task, it records adopted harnesses and supported lessons in a persistent improvement harness for future tasks. The README reports that a single step improved GLM-5 and Gemini 3.5 Flash on three professional benchmarks. It supports OpenCode and Codex executors and ships with three benchmarks: JobBench, LAB and WorkBuddy-Bench. The repository notes that it isn't an officially supported Google product.

Who it's for:Researchers and engineers who study or evaluate agents and can set up an isolated test environment. Readers who don't write code won't have a use for it.

Before you start: The README asks for Linux, Python 3.12 and uv. There's a local demo that needs no model credentials, so you can see how it works first. Running real model benchmarks also needs Node.js 20 or later, the Google Cloud CLI, Vertex AI access and billing, at your own cost, and task content, context and evaluation data are sent to the model endpoint you configure. The README says the executor's network access is unrestricted and Codex's own sandbox is turned off, and the code passes flags that bypass approvals and the sandbox. Some benchmarks start programs directly on the current machine with your own permissions. AI Inspiration World doesn't recommend running it on a computer with personal files, keys or signed-in accounts; use a separate virtual machine or container for research only. Setup also downloads upstream code and benchmark data. The code is licensed under Apache-2.0, and third-party data and components keep their own licenses.

Based on the project's public documentation. Not tested by AI Inspiration World.