To use ChatGPT agent mode, open the tools menu in the composer on a paid plan, select Agent mode (or type /agent), and give it one task with a clear finish line — it then opens its own virtual browser and works through the task, pausing to ask before anything consequential. The part most guides skip is the controls: what the permission prompts mean, when to use takeover mode, and how to stop a run cleanly.
Short answer: Turn on agent mode from the tools menu (paid plans only), describe one task with a clear deliverable, and watch the virtual browser window. Agent mode pauses automatically before logins, purchases, or anything else consequential, and asks you to confirm. Use "take over" for any login instead of typing a password into the chat, and you can stop the run at any point.
I run agent mode most weeks for price checks, research compiles, and first-draft spreadsheets, and in my testing the thing that trips people up is never the task itself — it's not knowing what the permission screen is asking, or forgetting takeover mode exists. This guide covers the settings and controls specifically: what to turn on, what each prompt means, and how to stay in charge of a task that's running on its own.
What you'll need
A paid ChatGPT plan — Plus, Pro, Team, or a business plan. Agent mode is not on the free tier; if that's you, my guide to what the free plan can still do covers what's available without paying. You'll also want a task with a real endpoint (a spreadsheet, a document, a specific answer), a few minutes to stay near your device for the first run, and login credentials ready if the task touches an account — you won't type them into the chat, but you'll need them for takeover mode. Nothing else to install; agent mode runs inside the same web app or desktop client you already use.
Step-by-step: How to use ChatGPT agent mode
1. Open the tools menu and select Agent mode
Start a new chat, click the sliders/plus icon in the composer, and choose Agent mode. Typing /agent does the same thing without touching the menu. If you don't see it, your plan doesn't include it yet, or your admin has it turned off on a business account.
2. Describe one task with a deliverable
Write the task as if you were handing it to a new hire: what to check, what to produce, and any rule you want followed ("flag anything unclear instead of guessing"). Vague missions wander; scoped tasks with a finish line get done. My companion guide on ChatGPT agent task framing goes deep on this part if you want more examples.
3. Watch the virtual browser window
Once you send the task, a small window shows the agent's screen live — you can see it clicking, scrolling, and typing. You're free to close the tab and come back later, but for your first two or three runs, watch a full task at least once so you understand its pace and habits.
4. Respond to permission prompts
Before anything consequential — submitting a form, making a purchase, sending a message on your behalf — the agent stops and asks you to confirm. Read what it's about to do, not just the button. Approve one action at a time rather than a blanket "always allow," at least until you trust the specific task type.
5. Use takeover mode for logins
When a task hits a login wall, click take over and sign in yourself inside the same browser window. According to OpenAI’s help documentation, the model doesn't see or store what you type during takeover — your password stays out of the chat entirely, which is the whole point of using it instead of pasting credentials into a message.
6. Stop, pause, or let it finish
You can interrupt a run at any point if it's off track, heading somewhere you didn't intend, or simply taking too long on a low-priority task. When it finishes on its own, it hands back whatever it built — a file, a table, a written summary — plus a rundown of what it did.
Agent mode vs. other ChatGPT modes
Agent mode isn't the only "do more than chat" option in ChatGPT, and mixing them up wastes a task. Here's how they actually differ:
| Mode | What it actually does | Where to find it | Best for |
|---|---|---|---|
| Regular chat | Answers in text only — no browsing, no file creation | Default composer | Quick questions, drafting, brainstorming |
| Agent mode | Opens a virtual browser, clicks and fills forms, builds files, pauses for permission | Tools menu → Agent mode, or /agent |
Multi-step research or web tasks with a deliverable |
| Deep research | Reads many sources and writes one long report — no clicking or form-filling | Tools menu → Deep research | A single thorough writeup, not an action |
| Tasks (scheduled) | Repeats a simple prompt on a schedule, no browsing | Tools menu → Tasks | Recurring reminders and simple recurring checks |
If your job is "go find out and build me something," agent mode is the right tool. If it's "just answer this well," regular chat or deep research does it faster.
Example prompts you can copy
These lean on the controls covered above — permission checkpoints and takeover — rather than just the task itself:
- "Turn on agent mode and check the pricing pages of [3 tools]. Ask for my permission before opening any page that requires a login."
- "Go through my most recent [connector] messages and draft replies to the simple yes/no ones. Save drafts only — do not send anything."
- "Check [product page URL] and tell me if the price or stock status changed since the last time you looked. Do not add anything to a cart."
- "Start this task, then pause and let me take over the login for [site] before you continue."
- "When you finish, list every step where you had to guess instead of finding a clear answer."
Each one tells the agent exactly where the boundary is — what it can decide on its own and what needs you.
Common mistakes to avoid
The one I see most: tapping "always allow" on a whole category of actions just to stop the interruptions, instead of reviewing each consequential step as it comes up — that's the setting most likely to turn a minor mistake into an expensive one. Second, walking away entirely during a task that touches money or a real account; stay reachable, even if you're not watching the screen. Third, typing a password straight into the chat box instead of clicking take over — the agent doesn't need to see it, and in my testing take over adds maybe ten seconds. Fourth, starting a long task without checking your remaining agent messages first, then running out mid-task and losing the progress. Fifth, skipping the end-of-run summary — a wrong guess buried in step six of ten is easy to miss unless you actually read what it flagged.
Tools that make this easier
Agent mode is strongest at boring, structured, web-based work — if your task is heavier on writing than browsing, a dedicated tool usually wins; my ranked best AI writing tools guide covers the ones worth paying for. If you're comparing ChatGPT's approach against Anthropic's, how to use Claude AI walks through the equivalent setup on that side, including Claude's own project-based memory. Coding tasks are a different category entirely — see the fuller field in ChatGPT alternatives for coding before you point an agent at your codebase. And if you're setting this up for a classroom, the tools in best AI tools for teachers solve most school tasks more directly than a general-purpose agent.
My take on agent mode's controls
The controls are the actual feature here, more than the browsing itself — any tool can click buttons, but the pause-and-ask pattern is what makes it safe to hand off real tasks. Once I stopped blanket-approving categories and started reading each permission prompt, the number of "wait, why did it do that" moments dropped to close to zero. Learn the tools-menu shortcut, use takeover for every login, and read the end-of-run summary before you trust the output — that's the whole system, and it holds up under real use.
Frequently Asked Questions
Is ChatGPT agent mode free?
No. Agent mode requires a paid plan — Plus and up — and agent tasks are metered separately from regular chat messages, with their own monthly cap. Free-tier users get a capable chat model but no agent mode; see our guide to using ChatGPT for free for what's included at no cost.
How long does it take to learn how to use ChatGPT agent mode?
Turning it on takes seconds. Learning to read the permission prompts and use takeover mode confidently takes one or two real tasks — most people get comfortable after watching a single run start to finish.
What is the easiest way to use ChatGPT agent mode?
Open the tools menu, pick Agent mode, and give it one task with a clear deliverable and a rule like "ask before anything you're unsure about." Watch the first run so you know what the permission prompts look like before you start trusting it unattended.
Can I stop a ChatGPT agent task once it's running?
Yes. You can interrupt the run at any point, take over the browser to do a step yourself, or just close the task if it's off track. Nothing keeps running once you've stopped it.
Is ChatGPT agent mode safe to use?
Reasonably, if you use the controls as intended: confirm each consequential action instead of blanket-approving, use takeover mode for logins instead of sharing credentials in chat, and treat any number it brings back as something to verify, the same way you'd check any web research.