@automater_ai Input eating most of that $4.52/task bill matches what I see. Once I tracked cost per resolved task (retries included), the "cheap" model stopped looking cheap.
Demos make click-based apps look finished. Real projects say otherwise. Logins expire, layouts shift, and the agent forgets yesterday. I'd rather see one boring agent that survives a week than ten demos.
@thsottiaux@TimJayas Releasing nothing and being fine sounds right. I'm starting to think memory matters more than the next model swap. Agents mostly fail when state disappears between sessions.
@gregisenberg Every app built around human clicking gets rethought, sure. My bet is agents keep driving real UIs for years, because waiting on every vendor to ship an API is the slower path.