Jev for computer use

A computer-use agent is a loop of irreversible actions — click, type, submit. Jev is the typed decision that sits inside that loop: before each action it returns a locked answer and a confidence score, so the agent only proceeds when it should.

Computer-use models are good at seeing a screen and proposing an action, and bad at knowing when not to. Jev is TypeSafe AI's System One model built for exactly that gap: a fast, calibrated decision that returns one of your options plus a confidence value, in roughly 70–500ms — fast enough to gate every step without stalling the loop.

Where Jev fits in the loop

• Safety gate: 'is this action reversible / safe to run?' → typed yes/no before it fires
• Action choice: pick one of the allowed UI actions instead of parsing free text
• Confidence stop: low confidence pauses the loop and asks a human
• Loop detection: flag when the agent is repeating itself and going nowhere

Why typed decisions beat a second prompt

You could ask a big model 'should I click this?' and parse the answer — but in a tight action loop that is slow, uncalibrated, and can return something you didn't plan for. Jev returns one key locked to your options with a calibrated probability from RLCD training, so the gate is deterministic in shape and about $0.001 per check. High confidence proceeds; the low-confidence tail routes to review.

Wire it in

Before each action, POST the proposed step and the screen state to https://jevtypesafeai.com/api/v1/decide with a jv_live_ key and read back the typed decision. Command Safety and the Agent Loop Detector are two computer-use gates already built this way.

See also: Command Safety · Agent Loop Detector · Jev for AI agents