Skip to main content
Abbie doesn’t produce every reply the same way. A quick factual question and a genuinely hard one go through different amounts of work, visible live as the run strip and recorded in full in the cognition log. The stages below are the full set she can draw on — not a fixed sequence every message runs end to end.
Works out what you’re actually asking. If your message bundles more than one distinct ask together, it’s split into parts here so each gets properly addressed rather than one crowding out the other.
Pulls in whatever memory and organisational context is relevant before drafting begins, so the answer starts from what Abbie already knows about you and the account rather than a blank slate. See Memory.
Decides how the request gets answered: as a single pass, split into independent pieces of work, or answered several different ways at once so the strongest can be picked. It also decides how many drafts that takes and whether the result needs an independent check afterwards. Your own Fast, Auto or Thorough choice steers this rather than overriding it outright — Fast skips the check, Thorough always runs one and compares more candidates, Auto leaves the call to the router.
Produces the draft or drafts the router decided on. A single task in Fast mode is one pass with nothing else after it; a request split into parts, or answered several ways for comparison, runs those independently before anything is judged between them.
When more than one draft was produced, an independent pass compares them and picks or combines the best answer. This step is skipped entirely when only one draft was made; there’s nothing to judge between.
A final, independent check of the chosen answer against your preferences and the facts available, before it’s sent to you. Skipped outright in Fast mode; always run in Thorough mode. See Confidence and verification.
The finished reply, delivered to the conversation. If it cited sources, this is also where those citations are checked and, if the answer used the web but didn’t come back with any, one repair pass adds them. See Citations.

​
Why a stage sometimes retries or falls back

If a model is briefly unavailable, Abbie retries automatically and, if needed, falls back to a different model to get you an answer rather than failing the turn outright. This shows as a retry count or a fallback badge on that stage in the cognition log; the answer still arrived, just not on the first attempt or the preferred model.