Skip to main content
While Abbie is working on a reply, a row of small pills appears above it: the stage strip. Each pill is a step in producing the answer, and it fills in live as that step finishes.

​
The stages

1

Classify

Abbie works out what you’re actually asking, including whether it’s more than one thing at once, in which case it gets split into parts before anything else happens.
2

Enrich

She pulls in relevant memory and organisational context before drafting anything, so the answer is grounded in what she already knows about you and the account.
3

Router

She decides how much firepower the request needs, and how many independent drafts to produce.
4

Experts

One or more drafts get produced in parallel. On straightforward requests this is a single pass; on more demanding ones several run side by side and the strongest is chosen.
5

Judge

When more than one draft was produced, an independent pass compares them and picks or combines the best answer.
6

Verify

A final check reviews the answer against your preferences and the facts available before it’s sent to you.
7

Output

The finished reply.
A stage pill can show as running, done, retried, or answered by a fallback model when the preferred one wasn’t available. Selecting a completed pill expands it to show which model handled it, how long it took, and, for a signed-in session, a link into the full cognition log for that step.

​
Collapsed and expanded

While a run is fresh, the strip shows every stage in full. Once you’ve moved on to read older messages, it collapses to a one-line summary: how long the reply took, how many experts ran, and whether verification passed cleanly. Expand it again any time to see the detail.
If a run is taking a while, that’s usually a harder request working through several draft attempts rather than something stuck. See Confidence and verification for what each stage is actually checking.