How long a turn runs

A turn has a step budget: 40 by default, where each model reply is one step. Near the end, the agent is told to wrap up. It finishes the edit in hand or summarizes what's left instead of being cut off, and you say "continue" to pick it up.

Productive turns keep going

A turn that's plainly making progress gets another round of steps instead of the wrap-up, and the transcript says so. It counts as productive when, among its recent steps, either:

• several edited files successfully, or
• most turned up something new (a file it hadn't read, a search or command it hadn't run)

and it hasn't been caught looping. Going back over calls it already made doesn't count.

This continues up to a hard limit (200 steps by default), so a big job such as scaffolding a new project, or a long debugging session, finishes in one go. Stopping the run works as usual at any point.

Settings: Settings > Agent Behavior > Loop control has the per-turn step count, the keep-going-while-productive switch, and the hard limit. If you had already set your own step count, it stays a hard stop until you turn the switch on.

Long replies

A single reply also has an output limit.

• When a reply is cut off there, the agent picks it up exactly where it stopped (up to three times in a row) instead of leaving a half-finished answer. The transcript notes each time.
• A tool call the limit cut off is never run on partial input. It's answered as "not run", and the agent redoes the work in smaller pieces.
• Either way, later replies in the turn ask for the model's full output allowance where Houston knows it.