Harness²

Google Research has put out Harness², a plug-and-play framework for LLM agents that is meant to help them evolve the way they work on their own. The pitch is simple enough: instead of freezing an agent’s setup, the system lets it keep refining the harness around it as it tackles open-world tasks.

That harness includes the usual moving parts — prompts, tools, and memory — but the twist is recursion. Harness² uses what it calls dual recursion: one loop for task-level Propose-Probe-Compose refinement, and another for accumulating reusable experience at the harness level. In plain English, it is trying to let agents learn from how they work, not just from the task outcome.

The headline number is up to 11.5 points better on professional benchmarks such as WorkBuddy-Bench. That is a serious-looking gain, especially in a space where lots of agent ideas sound smarter than they actually are once they leave the demo.

The other notable detail is the lack of ground-truth verifiers. That matters because open-world tasks rarely come with neat answer keys, and a lot of agent systems still depend on them more than they admit. Harness² is aimed right at that messier reality.