Steelman Labs.

We are building general UI perception and action for frontier models — the foundation reliable, fast and low-cost computer use will stand on.

Frontier models are strong enough to plan a computer-use task with surprising clarity. Watch them try to carry that plan out and the gap is jarring. By the time the decision is made, reality has already shifted — the page finished loading, a pop-up appeared, the authorization timed out. And even for something as trivial as scrolling a page, the screenshot → reason → tool-call loop is absurdly long and expensive. Planning has effectively been solved; execution has not. That gap is what stands between computer use and wide adoption, and it is the problem Steelman Labs exists to close.

The industry’s current answer is the harness — orchestration wrapped around the model that retries failed steps, parses screens, and stitches tool calls together. Harnesses make demos work; they do not make models more capable, and they cannot, by construction, be general — the moment an agent leans on per-app glue, it works precisely where the glue was written and nowhere else. The harness keeps growing because the substrate underneath is wrong: flat screenshots and serialized DOM throw away most of what makes a UI legible — structure, affordance, state, change. No amount of scaling on top of an impoverished representation will produce fluent computer use.

That is what Steelman Labs is for. We are working on general UI perception and action for frontier models. Put that foundation in the hands of builders everywhere, and the next generation of agents and end-user products can assume a model that actually sees a UI and actually acts inside it. It is infrastructure — and we are building it for the people who will build everything else.