Skip to main content
Submitting a prompt doesn’t generate an answer directly — it produces a plan you review first, then a live-running task you can watch.
1

Review the plan

After you describe what you want (e.g. in Deep Research mode), the agent proposes a numbered plan — its research angles, an analysis step, and a final write-up step — with an ETA.
A step-by-step research plan card with Edit Plan and Start Research buttons
There’s no drag-and-drop plan editor. Click Edit Plan and describe what you want changed in plain language (different ranking criteria, a dropped angle, an added axis) — the agent replies conversationally and regenerates the whole plan.
2

Start it and watch it work

Click Start Research. The view splits into two panes: your chat collapses to a short “Working…” status on the left, while the right-hand canvas streams a timestamped log — search angles, result counts, which sources it’s reading, and interim takeaways.
Live two-pane progress view with a timestamped research log on the right and a collapsed chat pane on the left
3

Get the finished document

Once research wraps up, the right pane switches into a Google-Docs-style editor titled after your topic, with Contents, Download, and Share controls and a live word count. It’s a fully editable rich-text document, not a read-only report — you can keep working in it directly.

​
A real failure mode to watch for

In testing, a task reported completion (“Done — the report is in your chat”) while the actual document contained only a Sources heading — the synthesized write-up itself was silently missing, despite the chat log showing sources had been read and a report was being written. Separately, the stated “4-5 minutes” estimate isn’t a reliable ceiling — a research run can still be mid-flight well past that.
Always open and check the actual document before trusting a “Done” status or relying on it for anything time-sensitive. A completed status is not proof the content is actually there.

​
Run modes

There’s no separate run-mode pill on the chat prompt bar anymore. Two controls shape how a chat task runs:
  • Model tier: the dropdown next to the send button (Flash, Pro, or Ultra). See Model tier.
  • Ask permission: a toggle in the + menu of the prompt box. Off (the default), the agent runs without stopping. On, it’s meant to ask before creating or editing artifacts. See Ask permission.
In earlier testing, the approval setting did not actually pause a full research run at any point, including at the report-writing step. Treat every run as if it will continue without stopping.

​
Startup modes

Startup, the website builder, has its own prompt bar with two different controls.
Startup prompt box with the Build menu open, showing Plan (Discuss before building) and Build (Make changes directly)
  • Plan: discuss before building.
  • Build: make changes directly.
Startup prompt box with the Eco menu open, showing Eco, Balanced, and Max, with Max locked
  • Eco: budget-friendly runs that take a bit longer and cost less.
  • Balanced: a steady mix of speed and quality for everyday builds.
  • Max: not available on every account. On accounts without it, it shows a lock and “Max is not available on this account.”