A message comes in. The agent decides. A reply goes out.
We closed the loop end-to-end. An inbound email now lands in a journey, the agent classifies it, drafts a grounded reply, and either sends it (when confidence is high) or routes it to review with full context. Every human edit feeds back as a learning signal. This is the smallest version of the product that's worth talking about.
-
New
Classifier + draft pipeline. A real prompted-LLM classifier decides per message: auto-send, route to review, or draft for human. Decisions go through a gate chain so the rules are inspectable, not a vibe score.
-
New
Knowledge grounding. Point the agent at your help docs (markdown for now). It chunks them with structure intact — headings, code blocks, lists — and grounds its replies in the actual sources it used, not a paraphrase.
-
New
Review queue. Drafts the agent isn't confident in land in a queue with the journey context attached. A human approves, edits, or takes over. We capture the draft-vs-sent diff so the system learns from every correction.
-
New
Real outbound. Approved replies actually send (we're using Resend on the wire). Sender identity, threading, and the round-trip back into the journey all work.
-
Improved
Auto-send routing. When the classifier decision is auto_send and confidence clears the threshold, the agent ships without a review hop. Conservative defaults — most teams will keep it review-only for the first week.