Run and review your first task
Write a clear prompt, follow each step across files, browser, terminal and Git, handle waiting-for-input, and iterate to a result you're happy with.
A task is one unit of real work you hand to an agent. Get this loop right — prompt, watch, answer, iterate — and you can delegate almost anything. This page walks through a full task end to end.
Write a clear task prompt
Agents do best with a clear outcome, not a vague wish. A strong prompt usually names three things:
- The goal — what "done" looks like.
- The context — files, links, or constraints the agent should use.
- The format — a doc, a table, an email, code, a published page.
Weak: "Help with our pricing."
Strong: "Write a pricing page with three tiers (Starter, Pro, Team), a feature-comparison table, and a clear CTA. Match the tone of the homepage. Output it as a draft I can review."
You can attach files for the agent to use as context. Keep one task focused on one outcome — start a new task for unrelated work.
Follow the live steps
Once you send the prompt, the agent goes to work in its cloud computer and you can watch each step as it happens:
- Files — it reads inputs and writes drafts, documents, or code.
- Browser — it opens pages to research or verify facts.
- Terminal — it runs commands when a task needs them.
- Git — it tracks changes so work is versioned and reviewable.
Each step is shown in the session timeline, so you always know what the agent did and why.
Handle waiting-for-input
Sometimes an agent needs a decision only you can make — a choice between options, a missing detail, or permission to proceed. When that happens, the task pauses and asks instead of guessing.
Just reply in the chat to unblock it, and it picks up exactly where it left off with full context. There's no penalty for stepping away: a paused task waits for you.
Long task running? You can leave and come back. Execution continues in the cloud, and the session is right where you left it.
Review and iterate
When the task completes, review the output in the session. Open any file it produced, check the steps it took, and decide what's next:
- Good enough? Use it, publish it, or download it.
- Almost there? Reply with specific feedback — "tighten the intro and add a comparison row for support" — and the agent revises in place.
- Worth repeating? Turn a task you'll run again into a reusable Skill or a scheduled Automation.
Iteration keeps the full history, so you can always see how the deliverable evolved.
Tips for better results
- One outcome per task. Split big asks into separate tasks you can review independently.
- Show, don't just tell. Attach examples or reference files when format matters.
- Front-load constraints. Tone, length, and audience are easier to set up front than to fix after.
Related
Quickstart — hire your first agent
Go from sign-up to your first deliverable in about five minutes — recruit an agent, give it a task in plain language, and review the result.
Create an account
Sign up with email, GitHub, or Google, verify your address, and land in your first workspace — no install required.