How to Use ChatGPT Agent Mode

0:16 To hand ChatGPT a task it carries out on its own, switch to Work (on the web, the app or the desktop app), or in a regular chat open the tools menu and pick agent mode or type /agent. Then describe the goal, the sources to use and when to stop for your review; the agent works in its own browser and files, pauses to ask questions, and asks before "high-impact actions". Both need a paid plan, apart from limited Work in the desktop app on Free and Go, as of Nov 6, 2026. What each is and costs is on Agent Mode and ChatGPT Work; this page is the how-to.

Quick facts

Work or agent mode: which one?

The two overlap. Agent mode arrived on Jul 17, 2025, when Operator was folded into it. ChatGPT Work arrived on Jul 9, 2026 as "an agent for longer, more involved tasks" that produces finished documents, spreadsheets, presentations, reports and Sites. OpenAI's agent help article still describes agent mode as a tool in Chat and does not mention Work, and no OpenAI page read on Nov 6, 2026 says plainly how the two will relate from here. OpenAI's status page still lists both "ChatGPT Work" and "Agent" as separate components. In practice: use Work for a deliverable (a deck, an analysis, a recurring report), and agent mode for a one-off web errand from inside a chat.

How to start a ChatGPT Work task

1. Open ChatGPT. On the web or phone, choose Work from the mode dropdown; in the desktop app, choose ChatGPT from the switcher, then Work from the toggle (OpenAI Help Center).
2. In the desktop app, choose local or cloud in the composer's options. Local can use your files, apps and the built-in browser with permission; cloud keeps running when your computer sleeps (OpenAI docs).
3. Pick a model if you want to: GPT-6 Astra for demanding work, GPT-6.1 Sol for complex work at lower cost, Luna for simpler tasks.
4. Write the goal (next section) and send it.
5. Follow along: Work shows its progress, and you can answer its questions, change direction or approve actions as it runs.

2:01 OpenAI's tip for a first task: pick one with "a clear outcome, a few source materials, and an output you can review". OpenAI's own getting-started video picks one like that: find a restaurant's email in Gmail and the team's headcount in Slack, then draft a reply that is reviewed before anything is sent.

How to write the goal

OpenAI's Work guide lists what a good request contains (OpenAI docs):

• The outcome: "a 10-slide deck comparing our three suppliers on price, delivery time and defects".
• The sources and plugins: type @ and a plugin name to point it at a tool, such as a connected drive or Gmail.
• Constraints: "use only these sources", "compare the top five options", a budget, a deadline.
• What success looks like: the file type, length, who reads it.
• When to stop: "stop before sending anything" or "show me a draft first".

The agent help article adds a warning: avoid vague prompts such as "handle everything", which give the agent room to act in ways you did not mean. More examples are on the ChatGPT Prompting Guide.

Sign-ins and taking over the browser

Some tasks need you to be signed in to a site. In agent mode, the agent pauses and you choose ••• > Take over browser to type the password yourself; "no screenshots are captured" while you are in control. Afterward it tries to continue, though it may need to restart the run.

ChatGPT Work's browser on the web and mobile can handle sign-ins too, since Aug 25, 2026: it shows the login screen, works with password managers, and "ChatGPT cannot see your username or password". This is for Plus and Pro plans. To clear what it stored, go to Settings > Cloud browser > Browser data.

0:04 In the desktop app Work has two more routes for sites and apps without a plugin, as OpenAI's video on them shows: computer use works desktop apps the way you do, and Chrome use works in your own Chrome, on websites where you are already signed in. OpenAI says both are only in the Work tab of the desktop app, not on the web or your phone, and that ChatGPT asks permission before using an app or website, once or remembered for next time.

Approvals, stopping and safety

• Confirmations. Work asks before consequential actions such as a reservation or a payment; agent mode asks before high-impact actions. In the desktop app, Enter approves a request and Esc declines it (OpenAI docs).
• Stopping. You can interrupt or redirect a running task at any point; OpenAI advises stopping at once "if something seems suspicious".
• Prompt injection. A web page or email can carry hidden instructions meant to trick the agent. OpenAI's example is a comment trying to get the agent to fetch a password reset code from Gmail. Its advice: enable only the apps you need, keep passwords and private details out of messages, and clear the remote browser's data after sensitive sessions (OpenAI Help Center).
• Safety reviews. A Work task may pause for a safety review; read the notice and its findings before continuing (OpenAI docs).
• Lockdown Mode turns agent mode off entirely (OpenAI Help Center). See Account Security.

Scheduling a task to repeat

• Agent mode: when a task finishes, select the clock icon to repeat it daily, weekly or monthly, and manage schedules at chatgpt.com/schedules. Each scheduled run counts toward your monthly limit.
• Work: ask for a recurring update, or use Scheduled Tasks. Tasks can run once, repeat, monitor for changes or, on Plus and above, react to events such as a new Gmail message, a Slack message or GitHub pull request activity. Cloud is the better choice for these, since your computer does not need to be awake.

Active scheduled tasks are capped at 3 on Free and Go, 5 on Plus, 10 on Business and Edu, and 15 on Pro and Enterprise, as of Nov 6, 2026.

Limits and usage

Agent mode allows 40 messages a month on Plus, 400 on Pro and 40 on Business and Enterprise; "Only initial user-initiated agent requests count", not clarifications or sign-ins. Work follows the same usage structure as Codex: an allowance per five hours (none on Pro), weekly limits that may also apply, then credits. On the web, your remaining Work usage shows in the sidebar. Longer tasks use more, so check before starting a big one. Numbers per model are on ChatGPT Usage Limits.

Frequently asked questions

How do I turn on agent mode in ChatGPT? In a chat, open the tools menu in the message box and select agent mode, or type /agent. It is available on paid plans on the web, iOS, Android and the macOS and Windows apps.

Is ChatGPT agent mode free? No. Agent mode is for paid plans only, as of Nov 6, 2026. Free and Go users get limited ChatGPT Work in the desktop app only.

Can ChatGPT's agent log in to websites for me? You log in, not the agent: take over the browser (••• > Take over browser) and type your password, during which no screenshots are captured. ChatGPT Work's browser on Plus and Pro shows the site's login screen, and ChatGPT cannot see your username or password.

Will ChatGPT's agent buy things without asking? It is designed to ask for confirmation before high-impact actions such as a payment or reservation. You can also tell it in your prompt to stop before sending or buying anything.

How do I stop a ChatGPT agent task? Interrupt it from the conversation while it runs; you can also redirect it with a new message. Delete or pause recurring runs at chatgpt.com/schedules or on the Scheduled page.

How many agent tasks do I get on ChatGPT Plus? Agent mode on Plus allows 40 messages a month, and Pro 400, counting only the requests you start. ChatGPT Work uses a separate allowance shared with Codex.

Sources

• ChatGPT agent, OpenAI Help Center, read Nov 6, 2026
• ChatGPT Work and Codex, OpenAI Help Center, read Nov 6, 2026
• Get started with ChatGPT Work, OpenAI docs (learn.chatgpt.com), read Nov 6, 2026
• Pricing, OpenAI docs (learn.chatgpt.com), read Nov 6, 2026
• Commands, OpenAI docs (learn.chatgpt.com), read Nov 6, 2026
• ChatGPT release notes, OpenAI Help Center, read Nov 6, 2026
• Scheduled tasks in ChatGPT, OpenAI Help Center, read Nov 6, 2026
• Lockdown Mode, OpenAI Help Center, read Nov 6, 2026
• Introducing ChatGPT agent, OpenAI, Jul 17, 2025
• ChatGPT pricing, OpenAI, read Nov 6, 2026
• OpenAI Status, OpenAI, read Nov 6, 2026
• Get started with ChatGPT Work, OpenAI, Jul 9, 2026 (video)
• Getting Started with ChatGPT Work, OpenAI, Aug 28, 2026 (video)
• Use Your Computer and Browser, OpenAI, Aug 28, 2026 (video)