How Cua Replaced Full LLM Calls with Fast Decision Models
Source
Cua
Author
Cua
Date
Key takeaways · AI-distilled
In Cua's jev-use recipe, app code builds concrete candidate actions with IDs from the current browser state. Hosted Jev only picks an ID, so it cannot invent tools, coordinates or arguments; the app validates the choice, executes it via Cua Driver, then re-observes.
Jev and CUA-S1-FORMS are text-only, so screenshot-only interfaces need a perception step such as Microsoft's OmniParser to turn pixels into labeled elements. Cua notes a missed control or bad label can spoil the decision before scoring starts, and parsing adds latency.
CUA-S1-FORMS has about 706K trainable parameters (a roughly 2.8 MB checkpoint) and scores each form element independently as use value, check, click or skip, given values already extracted. Cua calls it an English-centric research prototype; source, weights and data are MIT.
In an initial narrow form test, the specialist reached 99.7% decision accuracy versus 83.6% for hosted Jev, which was not fine-tuned for the task. Local scoring took 7-9 ms versus 260-280 ms per hosted call, though Cua says those timings measure different boundaries.
Cua warns a model score is not proof an action worked, since calibrationHow well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.Full definition → describes behavior across many examples. The application must still check fresh state, and in the handoff Cua wants to explore, an unexpected dialog or failed check returns control to a general LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → planner.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
Why it matters
Lays out a concrete, reusable architecture (observe, build candidates, score with a specialist, execute, verify) for replacing expensive full-LLM tool calls with cheap bounded decisions in any agentic loopThe cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.Full definition →.
