Your First Agent: Run One, Small, First
The previous article covered the kinds of agents. Here we guide you to actually pick one and run just one low-risk task. Screens and names differ by product, but the flow is mostly common. The key: don't delegate important work right away.
What to Prepare
- An account for the AI you'll use (see the previous "account creation" article if needed)
- Many agent features are on paid/upper plans. A payment method (pricing/terms change—check official)
- One low-risk task to try (pick from below)
The Common Flow (Nearly the Same for Any Agent)
- Enable: turn on the agent / Cowork / Computer Use feature in the service
- Grant permissions: if asked for browser/PC control, allow only the needed scope (an OS permission dialog may appear)
- Throw one task: start with a reversible task like below
- Watch it: execution is shown; if it looks wrong, stop it immediately with "Stop"
- Check the result: inspect the returned deliverable (text, file, screenshot)
Recommended First Tasks (Low Risk)
- Web research: "research XX on 3 sites and bullet the key points"
- File organizing: "sort the PDFs in Downloads by name (do not delete)"
- Information gathering: "5 of today's AI-related news with sources"
- Table creation: "make a comparison table of these 3 companies (don't send or save)"
Tasks to Avoid at First
- Email sending / social posting (irreversible)
- Purchasing/paying for paid services
- Operations involving delete/overwrite
- Processing confidential documents / personal data
Do these only after you're used to inserting "a human approves before execution." Your first run should be zero-risk.
Sticking Points and Fixes
- Permissions needed: PC-control types may need OS permission (e.g., macOS Accessibility)
- Login wall: it stops on sites needing a password. Try in a pre-logged-in browser/environment
- 2-factor auth: AI can't pass SMS auth. Pass it beforehand, or try within an auth-free scope
- Cost: metered ones rack up on long tasks. Check the cap setting (latest on official)
- Loses the goal partway: long sequences are a weak point. Re-throw in small segments
Once It Works
One successful run is enough to get the feel. Next, polish how to ask. The next article "Basics of Asking an Agent" covers what to delegate vs. do yourself, and how to instruct so it works as expected.
