Google's Harness²: an agent that improves its own set-up for each task

Google Research open-sourced Harness², where an agent proposes changes to its own prompts, tools, memory and skills for each task, tests them and keeps what works, without retraining the model. Google reports gains of up to 11.5 points on a workplace benchmark (WorkBuddy-Bench); those are its own results.

Why it mattersThe lesson for anyone using agents: the instructions, tools and memory around a model often matter as much as the model itself, so improve your set-up before paying for a bigger model.

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.