Skip to main content
Coming soon. Not in this release. This page says what the feature is, so that the shape is public before the code is.
Ask a fast model the small questions and keep the deliberate seat for the ones that need it. Most decisions in a run are small. Is this ticket the kind the policy answers? Has this note changed since the reviewer read it? Is the reply ready to send? Asking a full coding agent costs a turn of inference for each one, which is why gates like these are usually left out of a workflow rather than written into it. A fast decision model answers that kind of question with a typed choice and a confidence. A step will ask it first, and hand the work to the deliberate seat only when the confidence falls under the gate, or when the decision itself says the question needs thinking about. The Jev engine answers typed questions today; it doesn’t yet route work to a deliberate seat on its own. The point is what becomes affordable. A typed call can run on every item. A full agent turn can’t, so it ends up running on none of them.

​
Next steps