Turn
Giliran
One user message plus everything the agent does in response to it.
A turn is one user message and the entire response it produces: the model's reasoning, any tool calls, the tool results, and the final answer. It is the unit you pay for and the unit of conversation.
Turns are not the same as messages. A user message can trigger five tool calls and ten model calls, all inside one turn, and it is one turn in your bill. This is why a task that loops badly is expensive: each retry is another round of tool calls inside the same turn.
A useful metric for a working agent is turns per task, not tokens per task. It is a better signal of whether the agent is working efficiently, and it is much easier to improve: fewer turns usually means better tool descriptions and clearer instructions.
Related terms
Agent Loop
The repeated cycle of decide, act, observe. Where autonomy actually comes from.
Session
One bounded run of work with an agent, from the first message to the end.
Tool Call
The model's request naming a tool and its arguments. Structured output, not a side effect.
Cost per Token
What you pay for each unit of text the model reads or writes.
Latency
How long you wait. The time to the first token matters more than the total.
Want this working inside your team?
It starts with one conversation about your workflow.