What you see
Four tools produce rows in the conversation, and each persists as its own row:
Listing the conversation returns only these user-visible kinds. Plain assistant text outside them is dropped, not rewritten into a canned message — so a bot that has nothing to say says nothing.
Parked steps show as approval widgets; see Approvals.
When a turn stops
There is no round counter. Bot does not inherit Chat’s eight-round budget — a computer task legitimately takes many steps.
Background tasks and subagents
task runs work in the background and reports task_started / task_progress / task_completed or task_failed. Subagents cannot send_to_user and cannot nest another task; everything flows back through the parent bot, which decides what to tell you. See Tools.
Waking the guest
An active turn on a bot whose Cloud guest has hibernated resumes the guest before the model is prompted. The bot always starts a turn with its shell and desktop available; the prompt never tells the model to apologise for missing a computer. See Computer.Computer use, precisely
The planner describes the control it wants — the Save button in the dialog — and a second inference call returns framebuffer coordinates forgrounded_click. The planner never guesses raw pixels. Screenshots feed the next step.
Compared with Chat and Code
See How Chat works and How Code works.

