Three refinements to how the AI support agent handles conversations:
Faster first token, cheaper turns
We restructured the agent's internal prompt so the stable parts — your workspace rules, channel context, client details — stay identical across every turn of a thread, and only the genuinely per-message material varies at the end. That lets the model cache the shared prefix instead of reprocessing it on every customer message, which means faster time to first word on the highest-volume call in the product, at lower cost.
No more blank replies
Occasionally a turn could end with the agent having done its background work — looked something up, updated a record — but never actually written a reply, leaving the customer a blank message. The agent now detects that case and runs one repair pass to recover the response; if a reply genuinely can't be produced, the conversation routes to a human through the existing escalation path instead of going silent.
Long threads keep their memory
Previously, anything beyond the last 30 messages of a thread was represented by a one-line summary. The agent now maintains a rolling memory of the conversation that is folded forward incrementally as the thread grows, so history and recent messages stay contiguous. On a thread that spans weeks, the agent remembers what was discussed at the start — and it records what customers said as claims rather than established facts, so the memory can't be used to smuggle in false context.
All three are live for every workspace — nothing to do on your end.