Instead of a single answer — a supervisor that plans the work, hands tasks to specialized agents and shows progress as a task list.
For complex questions one model isn't enough: you need to crawl the site, look at search results, check indexing and pull it all together. Agent mode does that. A supervisor plans the work, dispatches tasks to specialized agents — a search researcher and a site analyst — and assembles the result.
Progress is visible as it happens: a task list streams into the chat and updates as work completes. It collapses to a single line with a "done / in progress / queued" counter, and shows elapsed time when it finishes.
This mode replaced April's Research mode, which lasted a week: the same idea in design, but considerably more reliable in practice.
Broad tasks like "analyze my site", where you can't know in advance how many checks will be needed.
Turn on agent mode with the toggle in the input row and ask a broad question. The plan appears as a card above the answer — click it to expand the task list.