Riley Coyote wires Jev into an always-on computer use agent
An architecture post, not a benchmark, and so far the only amplification comes from TypeSafe AI.
Riley Coyote posted a description of an app he is building where an agent is always running alongside you, sees what you see, and uses your computer through Jev. In his split, Haiku writes and dictates text on screen because it is fast and cheap, and Opus 5.5 does the reasoning and acting, run through CC. Jev is the part that touches the machine, which he calls the agent's tendrils.
ambient intelligence, with @typesafeai's Jev as your agent's tendrils
That is the whole of what he published. There is no latency figure, no cost per decision, no task success rate, and no description of what the agent was asked to do. He does say the thing he told Claude while building it, that his primary goal was for the app to feel magical for the user, like it should not be possible. TypeSafe AI quote-posted that line back with the reply that with Jev it is possible.
What is actually new here
The pattern is the interesting bit, not the claim. Most of the Jev computer use work so far has put Jev in the control loop and let a bigger model sit out of the hot path. Riley Coyote is splitting three ways instead of two, a cheap model for the text that lands on screen, a frontier model for the reasoning, and Jev for the actions in between. Whether that split pays for itself is exactly the thing the post does not measure.
For people building on it
Treat this as a sketch of an architecture, not evidence that it works. The only numbers anyone has posted on Jev driving a desktop came from Milind S, who clocked about 90ms per decision on a Mac, and that was a different setup with a different stack underneath it. Nothing in Riley Coyote's post says what his loop costs, how often it picks the wrong action, or how it behaves when the screen changes under it, which is where always-on screen agents usually fall over.
The second thing worth noticing is who is amplifying. The reach on this one so far comes from the vendor, which is a claim about possibility rather than a result from someone who ran it. If you are deciding whether to put an ambient agent in front of a user, wait for the app, or for someone to post a trace with timings on it. Until then this is one builder saying he has something working on his own machine, which is a reasonable place to start and a bad place to copy from.
with Jev it's possible!
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.