Google Open-Sources Harness² for Self-Improving AI Agents

Building an AI agent usually means designing its setup once by hand, then watching every new task run through that same fixed setup. Google just open-sourced Harness², which lets the agent rework that setup for each task on its own.

The model weights never change. In Google's tests, a single round of edits was already enough to beat the original harnesses.

What Happened

A harness is everything wrapped around the model: the system prompt, tools, memory, skills, and subagents. Today a person designs it once, and the agent runs every new task with that same fixed setup.

Harness² hands that design work to the agent. On every task it proposes edits to its own harness, runs a few candidate versions, and composes a better one by comparing how those runs behaved. Because it learns from the differences between runs, it needs no correct answer or grader for the task.

It also carries lessons forward. After each task it saves the harness it settled on, along with the notes that held up across the runs, so the next round of edits starts from those notes instead of from scratch. Every finished task makes the next one more informed.

Google tested it with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. One round of edits beat their original harnesses, and more rounds pushed scores higher. Those extra rounds can go wide or deep: the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.

Why It Matters

Most agent tuning assumes you can measure the result, which means a labeled answer or a grader. Harness² skips that requirement, so tasks where nobody has written down the right answer can still produce a better setup over time.

It also changes what one good run is worth. Instead of a harness that stays frozen after launch, the agent keeps a record of what worked and starts the next task ahead of where it began. If you build agents, the repo is the place to look, and the idea is worth trying on whatever setup you already run.