Small decisions. Interesting possibilities.Submit contentSubmit

Handsy

An iPhone-control agent that uses Jev for the small, frequent steps and a larger model only for the rest.

Added to Jevfast

How it uses Jev

Inferred from the creator's launch post: before each step of an iPhone-control run, the agent asks Jev to settle the routine part of the step and leaves the rest to Qwen 3.8 Flash, so the expensive model is not called for every action. The post shows a full Duolingo check-in run performed this way and says TestFlight 1.0.3 carries the pairing, but it does not show the typed questions, the state each step sends, or the confidence threshold that escalates a step to the larger model; those remain unknown.

Why it is interesting

The cost lever sits inside the agent loop rather than around it: a decision model absorbs the repetitive steps that a device-driving agent repeats on every run, and the language model is spent only where judgement cannot be narrowed to a typed question.