Portals, rendered tables, paginated lists and PDFs behind a search form. The agent opens the real page, follows the pagination, and hands back CSV or JSON with a source URL on every row.
Unit prices instead of sticker prices, stock checked across twenty products at once, coupons tried one at a time at the real checkout. The agent stops before payment, every time.
Availability that only exists inside a live calendar widget, seat maps you have to read visually, walking distances you have to check on a map. These are the flows a scraper cannot see.
Regulators, clinical registries, academic publishers and city halls. Sources that are public but not queryable, where the answer is three clicks deep in a form nobody has an API for.
Directories, portfolio pages, exhibitor lists and job boards, turned into a deduplicated table with every claim traced back to the page it came from.
Four browsers at once, a task that runs every weekday at eight, a workflow saved as a reusable skill, a logged-in session that stops before anything irreversible.
Both platforms are hostile to automation and generous to a logged-in human. The agent is the second kind: it reads your feed, exports your own followers, drafts in the real composer, and stops before it posts.
Tetris, chess, GeoGuessr, powerline.io. Not a benchmark - a demonstration that the agent can read a board that changes faster than it can think and still act on it.
0:46
0:27
0:48