What is a node-based voice agent?

A single prompt is like briefing a new hire once and hoping they remember everything. A flow is like giving them a checklist with a page per stage of the call.

Each node is one stage: greet and confirm identity, explain the reason for the call, handle the answer, close. The node holds its own instructions, which can be long and specific. The links between nodes are conditions, written the way you’d explain them to a colleague: “the caller agrees to pay”, “the caller says it’s the wrong number”.

During the call, the model works inside the current node. When it decides a condition is true, the call moves on and the next node’s instructions take over. The conversation so far comes along, so the agent doesn’t forget what the caller already said.

Flows earn their keep on calls with real branches. Wrong person, reschedule, refusal and “speak to a human” are four different paths, and each needs its own rules. Squeezing all of that into one prompt makes the model lose its place, especially late in a long call.

You also get a map of what happened. Because each call passes through named nodes, you can see where callers drop off, which branch is most common, and what they said at each step. You can pull fields out per node too, like the promised payment date.

The cost is flexibility. A tightly drawn flow can feel rigid when a caller goes somewhere you didn’t plan. The usual fix is a global node with shared rules, and conditions loose enough to catch how people actually talk.

Start with a prompt. Move to a flow when the branches start fighting each other.