Multi-turn conversation
A multi-turn conversation is an exchange in which later questions, answers, or actions depend on information established in earlier turns. Unlike a single question and response, it requires the system to preserve context as the caller adds details, makes corrections, asks follow-up questions, or changes direction.
Most useful phone workflows are multi-turn. Booking an appointment may require a service type, location, date, and time gathered across several exchanges. A caller might then say “Make it the later one,” which only makes sense if the agent remembers the options it just offered. That dependency on prior dialogue is the defining feature.
Supporting multiple turns involves more than retaining a transcript. The agent must track which facts are still relevant, which values have been replaced, what action is in progress, and what remains unresolved. Dialogue management uses that state to select the next question or action. A context window gives the underlying model access to some portion of the instructions, conversation, retrieved material, and tool results, but access alone does not guarantee correct state tracking.
Phone calls add pressure because there is little room for visible correction. Pronouns, brief replies, interruptions, and background noise can make a turn ambiguous. If the caller says “yes,” the system needs to know which question is being answered. If the caller corrects a phone number digit, it should update the value rather than append another number or restart the intake.
Good multi-turn design keeps the conversation flexible without losing control of the task. The agent should acknowledge meaningful corrections, avoid asking for information already provided, summarize before consequential actions, and recover when the caller introduces a side question. It should also recognize when context is too uncertain and ask a focused clarification instead of guessing.
Evaluation should use complete call scenarios. Test callers who answer out of order, refer to earlier options, revise a key detail, pause mid-answer, switch goals, or return from a detour. A system that sounds strong in isolated samples may still fail when information must survive several turns. The practical standard is continuity: the right facts remain attached to the right task until the call reaches a valid outcome or an appropriate handoff.