How to Use ChatGPT Agent Mode

ChatGPT agent mode is OpenAI’s hosted agent: tell it what to do, it browses, clicks, reads, and comes back with a result. It’s the easiest way to run an agent today — no code, no SDK, no infrastructure. It lives inside a ChatGPT Plus, Pro, or Team conversation.

This guide walks you through enabling it, a first real task, and the guardrails you’ll want to set before you trust it with anything sensitive. Written April 2026 against the current ChatGPT agent experience.

What it is (and isn’t)

Agent mode lets ChatGPT use a browser, a terminal, and a file workspace to complete a multi-step task while you watch. Think: “find the three cheapest flights, compare on layover length, and write the comparison to a doc.” It is not a background worker — it runs live inside your conversation and stops when you close it. For always-on automations, see our setup overview.

Step 1 — Confirm you have access

Agent mode ships on Plus, Pro, and Team plans. Free-tier accounts don’t see it. If you’re on Enterprise or Education, your admin may have toggled it off; ask them first before you spend an afternoon debugging.

Step 2 — Turn it on inside a chat

Open any conversation. In the composer, click the tools dropdown and select agent mode, or type /agent to activate it. You’ll see the composer change — ChatGPT is now in agent posture rather than chat posture.

Step 3 — Write a good first prompt

Agent prompts aren’t chat prompts. They should include:

• Goal — the outcome you want.
• Sources — which sites or accounts to use.
• Stop condition — when you consider it done.
• Output shape — a doc, a table, an email draft.

Example: “Find three independent bookshops in Lisbon open Sundays, write their addresses and opening hours in a markdown table. Stop after three; don’t list more.” Use our prompt improver to tighten the prompt before you run — agent tokens cost more than chat tokens.

Step 4 — Watch the first run end to end

The first time, do not walk away. Watch it open tabs, click, and reason. You’ll see two things: where it’s smarter than you expected, and where it confidently does the wrong thing. You can interrupt with “stop and ask me” or “use this site instead” mid-run.

Step 5 — Approve sensitive actions manually

For anything that writes — sending email, posting on a site, clicking “buy” — agent mode will pause and ask you to approve. Don’t disable this. An agent that confidently sends the wrong email to your boss is a bad day. The small friction of one click is worth it.

Step 6 — Log in only when you need to

Agent mode can ask you to log into sites. Do it inline, in the agent’s browser window, for that single session. Don’t save passwords in the agent; use your password manager and type them in when asked. Never share two-factor codes or banking credentials with an agent — not this one, not any other.

Good tasks for agent mode

• Researching a short list (books, flights, apartments, tools).
• Comparing 3–10 items across a fixed set of criteria.
• Pulling data from a site that has no API.
• Drafting an email, doc, or spreadsheet from browsing.
• Light form-filling with your supervision.

Bad tasks for agent mode

• Anything that takes more than ~20 minutes end-to-end.
• Tasks that run overnight or on a schedule (use an SDK or Zapier).
• Large-scale data collection (hundreds of pages).
• Anything with your real money or real credentials.

When to graduate off it

If you find yourself running the same agent-mode task weekly, you’ve outgrown it. Move to an SDK or platform that can schedule the job — see the OpenAI Agents SDK guide for the developer path, or our agent setup overview for the full decision tree.