Cua: Fast Decision Models Instead of Full LLM Calls in Agents

Cua, which builds tools that let AI agents operate computers, is testing how many of an agent's decisions really need a large general-purpose model. In a long post, Cua argues that once an agent reaches a form, the remaining choices are local: does this value belong in this field, is that box already checked, should this element be left alone. Its proposal is for the application to list the allowed actions, let a decision model score them, and have ordinary code validate, execute and check the result, instead of asking a large language model to write out another tool call.

The starting point is TypeSafe AI's Jev, which TypeSafe calls a System One model. Jev reads text, including JSON describing the current screen, and returns a typed decision, such as one of the options it was given or a probability for a yes-or-no question, rather than free text. Cua's preview recipe, jev-use, connects hosted Jev to Cua Driver, its desktop automation tool: the application builds candidate actions with IDs, Jev picks an ID, and the application resolves it, validates it and executes it, so Jev cannot invent tools, coordinates or arguments. Neither Jev nor Cua's own model reads pixels. Cua sketches using a screen parser such as Microsoft's OmniParser to turn a screenshot into labelled elements first, and calls that an architecture to explore rather than an integration it is shipping.

Cua's own model, CUA-S1-FORMS, is a form specialist with 706,048 trainable parameters and a checkpoint of about 2.8 MB. Given form elements and values already extracted from a document, it scores each element as use this value, check, click or skip. In Cua's first head-to-head on that task it reached 99.7% decision accuracy against 83.6% for hosted Jev, and scored a form locally in 7 to 9 milliseconds against 260 to 280 milliseconds per hosted Jev call. Cua attaches the limits itself: the specialist was trained for exactly this task while Jev was not fine-tuned for it, the timings compare a computation on the local machine with a network round trip, and neither figure is a general computer-use benchmark or an end-to-end completion time. Cua says the source, weights and synthetic dataset are public under the MIT license, with both projects in its GitHub repository, and describes the model as an English-centric research prototype.

The larger plan, in which a general model plans unfamiliar steps and recovers from surprises while specialists take the familiar narrow ones, is presented as a direction, not a shipped system. Cua adds that knowing when a specialist is out of its depth is itself a hard problem that a high score does not answer. Intel's item on the CUA-S1 release covers the rest of the model family.