Turn

A turn is the unit of back-and-forth in a session: you say something, the agent works until it has nothing more to do without you, and control comes back. From your side that is one exchange. Underneath, a single turn can contain many model provider requests, because after every tool call the harness feeds the result to the model and asks what comes next. A turn that reads eight files and runs the tests twice is a dozen round trips billed as one conversation step.

This is the mismatch behind surprising bills and slow responses. The cost of a turn is not the length of your message; it is how many tool calls the agent needs and how much context gets resent on each of them. Late in a long session every request carries the whole transcript, so the same simple turn costs several times what it cost at the start.

Shorter, more decisive turns tend to work better than sprawling ones. Give the agent a bounded task, let the turn run, review, then take the next turn with what you learned. If a turn is heading somewhere wrong, interrupt it (see steering) rather than letting it complete a plan you already know is off.