Changelog

Auto picks its tier from the whole conversation

Auto now reads the conversation, not only your latest message, when it chooses between Fast, Balanced, and Max. A short reply such as "yes", "go", or "2" gets the tier of the work you are agreeing to, so saying yes to an app change no longer lands on Fast. On conversations replayed from the past month, Auto now picks the tier the work needed about 90% of the time, up from 30%, and sends work to a tier too light for it in under one case in ten, down from more than half. Read how Auto chooses on the models page.

How it works

  • The reply you answer counts. Auto reads the end of the previous reply alongside your message, so it knows what "do it" refers to.
  • Follow-ups can step up. Within a few minutes of the last message Auto keeps the same model, which keeps the conversation quick, but a follow-up that needs more, like turning a question into an app, now moves up a tier. It never steps down mid-conversation.
  • Files count. A message that is only an attachment is routed by what it carries instead of going straight to Max.
  • A new specialist's first turn runs on Balanced. It looks through your accounts and asks a few questions, which never needed Max.

Also shipped

  • New conversation now weighs the workflows, apps, and skills each specialist owns when it suggests where a message belongs, sends one-off errands to your Chief of Staff, and suggests a new specialist only when you ask for one or describe ongoing work nobody covers. On real workspaces we replayed, it picks the right specialist about 9 times in 10, up from 7, and stopped suggesting a new specialist for work one already owns.

Your first specialist starts here.

Give it one job. Build the workflows, apps, and skills to get it done.