Skip to main content
Two limits are worth knowing before you design a pathway. They are not knobs for you to tune — they are the safety net that pulls a conversation back when a flow misbehaves. Understanding them lets you predict how your agent behaves in edge cases.

​
The two limits

A “turn” is one round trip: the user says something, the agent replies. Within a single turn:
  • At most 4 node transitions (moves from one node to another)
  • At most 4 tool steps (calling several tools at once still counts as one step)
The two limits are independent: using up all four node transitions does not consume tool budget, and vice versa. When either limit is reached, the agent never goes silent — it always closes the turn with a spoken reply. Both limits reset every turn — they are not a total for the whole conversation. A long call moving through dozens of nodes is perfectly normal.

​
Moving through several nodes in one turn is intentional

The first time you see a turn cross three nodes, it can look like a broken flow. Usually it isn’t — it means the user said a lot in one breath. Picture a service-booking flow designed as four nodes: ask for the service → ask for the vehicle → ask for the mileage → confirm everything. If the user opens with:
“I’d like to book a routine service. It’s a 2019 model, around 80,000 kilometres.”
All three questions are already answered. The correct behaviour is to jump straight to the confirmation node, not to walk the node order and ask “What service would you like?”, “Which model?”, “What mileage?” all over again. Re-asking what the user just said sounds like the agent wasn’t listening — one of the most damaging habits in a voice conversation. So multi-node turns are a feature, not a defect. The limit exists only to keep them from running away.

​
Making the agent speak at every node

If your scenario genuinely needs step-by-step confirmation (reading each detail back so the user can verify it), that is a prompt-level requirement, not a flow-level one. How to do it:
  • Say so in the node prompt: “Read back the information collected so far and get confirmation before moving on.”
  • Add a precondition such as “the user has confirmed” to the outgoing edge condition, so moving on without a reply isn’t possible.
  • Step-by-step confirmation makes calls longer. Add it only to nodes that genuinely need verification, not to the whole pathway.
Do not put an instruction like “route to the next node without speaking” in your global prompt. That explicitly authorises the agent to chain transitions in complete silence — which the user experiences as an unexplained pause after they finish talking. “Don’t speak” and “speak at every step” are opposites, and it is easy to write the wrong one.

​
About cyclic pathways

If your pathway contains a cycle (A → B → A), the node transition limit is the last line of defence: the agent won’t loop forever, and once the limit is reached it stays put and finishes the turn. Don’t treat that as permission to design cycles loosely. A cycle without a clear exit condition will already have the user listening to the agent go in circles before the limit is reached. Give every loop-back edge an explicit condition (for example, “the information provided is still incomplete”) and make sure at least one path always leads out.