arc-cua hands bounded desktop subtasks to a fast decision model so the frontier model stops paying for every click

GitHub (shhivv/arc-cua)·medium signalXBlueskyLinkedInCopy link

Isle released arc-cua on 2026-09-20 (MIT, 107 stars in a day), a Python action layer where a planner emits a payload with a goal, inputs, verification conditions, constraints and a max_actions budget, and a decision model executes the UI loop and returns a status plus the action count. The planner deliberately lives outside the package, so any GPT, Claude, Gemini, local model or deterministic planner can drive it. This is the computer-use equivalent of the Jev routing pattern: the expensive model sets intent, a cheap model closes the loop.