Google Research just open-sourced a framework to autonmatical improve AI harnesses
by•
Google Research released Harness2, and I think it reiterates an important direction for AI-native applications.
The idea is that the model isn't necessarily what needs to improve. It's the harness around it: prompts, tools, memory, skills, and orchestration.
Harness² experiments with automatically improving that harness based on how agents perform on tasks, without retraining the underlying model.
As a founder building AI agents, this is a super manual task today.
Curious what other builders think. Will the harness become a bigger competitive advantage than the underlying model?
23 views
Replies
This is huge for builders... I'm still skeptical that it replaces actual model training though. At some point, you still have to fix the underlying issues in the weights and not just patch around them with better harnesses. Definitely still watching how this plays out!
From a small product's point of view the harness is already where the advantage lives, because the model is the one part I can't own. My app generates songs. Over a month I swapped the main music engine twice and kept a second one as a fallback, and the app didn't notice, because everything users actually feel sits around the model: a step that screens the request before it goes out, a step that recognises the engine's refusal and rewrites the prompt instead of showing an error, a step that listens to the finished track and realigns the lyrics to what was really sung, since singers drop and merge lines. None of that is in any model's weights, and all of it is what makes the product feel finished. Where I agree with Amanda: the harness patches around a weakness, it can't add an ability. If the engine can't sing in a language, no amount of orchestration fixes that. So my take is that the model sets the ceiling and the harness decides how close to it you get, and for most of us the gap between a raw model call and a good product is far bigger than the gap between this year's model and next year's. Automating the harness tuning is interesting, but the hard part in mine wasn't tuning, it was knowing what to measure. Nothing told me the lyrics drifted until I listened.