Dialogue state is the structured record of where the conversation currently is. Identified caller, declared intent, filled slots, pending confirmations, tool results received, history of turns. The agent reasons against the state at every turn; the state is what makes the conversation coherent across multiple exchanges.
In template-based systems, dialogue state was an explicit object with named fields. In LLM-driven systems, much of the state lives in the conversation history that gets fed back to the model. Either way, a serious vendor maintains an explicit state object alongside the LLM context, so the system can reason about 'has this caller confirmed their date of birth' without re-scanning the transcript every turn.
The state is also what makes handoffs clean. When the conversation transfers to a human, the human picks up the explicit state, who the caller is, what they want, what is filled, not a transcript they have to read. A vendor that lacks explicit state ships handoffs that feel like dropping a stranger into the middle of a meeting.
Vorel maintains a typed dialogue state per call. On handoff the human operator sees the state alongside the transcript, with confirmed and unconfirmed values clearly marked.