Ari should be your default first responder, and has been rebuilt from a rigid "classify and handoff" workflow into an agentic search first approach.

Rather than running a single fixed pipeline to answer your customer's questions, Ari now iterates: it first searches your knowledge sources, evaluates what came back, refines the query if needed, then drafts an answer. A judge can reject the answer, meaning Ari retries with feedback rather than immediately handing off, helping your customers get answers more often without escalating to a human.

Alongside this, a number of other improvements were shipped:

- Ari's question classification is stronger, meaning more of the right type of questions flow through to search and answer rather than being routed to handoff

- The judge can now trigger retries with feedback when it rejects an answer, so a borderline response can be improved rather than discarded

- Every tool inside Ari now reads from a single consistent view of the thread, fixing edge cases where Ari only saw partial context