Run and review tasks
Give an agent a task in plain language, watch it work step by step, answer when it needs you, and collect the results.
You direct; the agent ships. A task is a single run inside a session: you describe what you want in plain language, the agent executes it live in its cloud computer, and you review the output. You stay the CEO — you never have to do the work yourself, only point and review.

Dispatch and review a task
Open an agent and describe the work
Open an agent from the sidebar, type your request into the message box, and send it. Write it the way you'd brief a teammate — "Research our top 5 competitors and write a one-page summary," not a list of commands. For models that support it, pick the reasoning effort for this run before you send.
Watch the steps and artifacts live
The agent starts working immediately. You see each step as it happens — files it reads and writes, commands it runs in the terminal, pages it opens in the browser, commits it makes in Git. Deliverables (documents, images, code, a built site) appear as artifacts you can open and inspect without leaving the session.
Answer when the agent needs you
If the agent needs a decision or missing detail, the session moves to Needs input (waiting_for_input) and pauses. Reply in the same conversation to unblock it — the agent picks up exactly where it left off, keeping all its context. This is the natural pause/resume: work waits for you instead of failing or guessing.
Review and follow up
When the run finishes, the session shows as Done. Review the artifacts, then send a follow-up message in the same session to refine ("make it shorter," "add a chart") — the agent keeps the full context, so you iterate instead of re-explaining.
Manage your sessions
- Rename a session to keep your work organized — titles show up across the agent view and the sessions dashboard.
- Track status at a glance — every session carries a badge: Queued (
pending), Running (in_progress), Needs input (waiting_for_input), Done (completed), Failed (failed), or Cancelled (cancelled). - Run several at once — agents work in parallel, so you can dispatch tasks to multiple agents and check back on each.
Need to run the same task repeatedly? Turn it into a scheduled task so it runs on a cron, interval, or one-time schedule without you starting it.