Harness² from Google Research is a plug-and-play recursive agent harnessing framework that lets LLM agents automatically evolve their execution harnesses (prompts, tools, memory) for open-world tasks—without ground-truth verifiers. Dual recursion: task-level Propose-Probe-Compose refinement + harnessing-level accumulation of reusable experience. Up to +11.5 points on professional benchmarks like WorkBuddy-Bench.

A harness includes everything surrounding the model—such as the system prompt, tools, memory, skills, and subagents. Traditionally, this setup is designed once and remains unchanged for every task the agent performs.

Harness² introduces adaptability, enabling the agent to independently tailor its setup for each task while keeping model weights static.

With this approach, the agent proposes changes to its harness, tests multiple configurations, and refines it based on performance insights from those trials.

Because it learns from run-to-run variations, there's no need for a defined correct answer or an external grader to assess results.

This process isn't limited to single tasks. Harness² also transfers learnings across tasks.

Once a task is complete, the agent retains the optimized harness and the recurring insights from those runs. When working on a new task, it builds on this accumulated knowledge rather than starting over.

This continuous refinement enhances future task performance.

In trials with Gemini 3.5 Flash and GLM-5 on legal, office, and professional benchmarks, Google found that even one round of edits surpassed the performance of fixed harnesses, with further iterations driving additional improvement.

This points to self-healing, self-growing harnesses... 🤯

Rewriting its own harness between runs without a grader telling it what correct looks like is the part I find hard to believe, so I want to see whether the accumulated experience actually transfers between unrelated tasks.