How do you use ChatGPT agent for a multi-step task?

ChatGPT agent, released by OpenAI on July 17, 2025, lets ChatGPT work through a task in several steps instead of answering in one reply. It runs on its own virtual computer, with a visual browser for clicking through websites, a text browser for reading quickly, a terminal for running code, and access to connectors, which link ChatGPT to apps such as Gmail or GitHub. It can hand back editable spreadsheets and slide decks.

The skill that matters most is the same one that matters in ordinary prompting: say clearly what you want. The difference is that the agent will act on what you say for many minutes, so the limits you set at the start carry more weight.

Before you start

Agent mode is available to Pro, Plus and Team subscribers, rolling out over the days after launch. OpenAI gives Pro users 400 agent messages a month and other paid users 40, with extra usage available through credits. It is not yet available in the European Economic Area or Switzerland.

To turn it on, open the tools dropdown in the message box and select agent mode. You can do this at any point in a conversation, so you can discuss a task normally first and switch on the agent once the plan is clear.

A worked example: comparing tools in a spreadsheet

Suppose a volunteer group needs a shared calendar and task app, and you have been asked to compare options. This is a good agent job: it needs browsing many pricing pages, pulling the same details from each, and organizing them in a table. It needs no purchases and no sign-ins.

I volunteer for a community garden with about 30 members. We need one app for a shared calendar and task lists.

Task: compare five well-known options that offer a free plan or a nonprofit discount.
For each, find on the official website: name, free plan limits, price per user per month for the cheapest paid plan, nonprofit discount (yes/no, and how to apply), mobile app (yes/no), and the URL of the pricing page.

Rules:
- Use only official websites for prices; note the date you checked.
- Do not sign up, sign in, start a trial or enter any information on any site.
- If a price is unclear, write "unclear" rather than guessing.

Output: a spreadsheet with one row per app, then three sentences recommending one option and saying why.
Before you start, ask me anything you need to know.

What typically happens next, described here as an illustration:

1. It asks questions. For instance, whether members mostly use phones, or whether there is a budget ceiling. Answer briefly.
2. It works and narrates. An on-screen narration shows each step: searching, opening a pricing page, reading a table. You can watch or leave it running.
3. You steer. If it chooses a tool you have already ruled out, interrupt with a message such as "Skip that one, we tried it last year." OpenAI says it picks up where it left off without losing earlier progress.
4. You check progress. If it seems slow, you can pause it, ask for a progress summary, or stop it and take the partial results.
5. It delivers. You get the spreadsheet and the short recommendation. On the ChatGPT phone app, you get a notification when it finishes.

Then do your part: open three of the pricing links and compare them with the table. Prices and plan names change often, and the agent can make mistakes, which OpenAI states plainly about this early release.

Tips for better multi-step results

• Give it a template. You can upload an existing spreadsheet for the agent to edit or use as a model, so the columns match what you already use. That option is not yet available for slide decks, and OpenAI calls its slideshow output a beta.
• Keep connectors off unless needed. A comparison task does not need your email. OpenAI recommends disabling connectors a task does not use, to reduce exposure to malicious instructions hidden in web pages.
• Choose the right mode. Deep research is still available separately in the composer dropdown and may give a longer, more detailed report when you only need reading and writing, not actions.
• Repeat what works. A finished task can be scheduled to run again, for example a weekly report every Monday morning.

What should it ask you before acting?

Good agent work starts with questions, so invite them. Ending your prompt with "Before you start, ask me anything you need to know" lets the agent clear up scope, budget and preferences before it spends minutes on the wrong path. OpenAI notes the agent may also ask for details on its own when a task is unclear.

During the task, there are moments when it must ask. It requests your permission before actions with real-world consequences, such as making a purchase. Some critical tasks, such as sending email, require you to actively watch. When a site needs a sign-in, you take over the browser and type your credentials yourself; OpenAI says it does not collect what you enter in that mode. Never paste passwords or card numbers into the chat, and do not approve a step you did not ask for, especially if a web page seems to have suggested it.

If you want stricter control, write your own stop points into the prompt: "Stop and ask me before you sign in anywhere, add anything to a cart, or submit any form." The agent then pauses at those moments, and the decisions stay with you.