ChatGPT agent, the "agent mode" you started from the tools menu or with /agent, is no longer available. OpenAI's help center now says to use ChatGPT Work for longer, multi-step tasks and finished deliverables. Work does the same kind of job (browse, use your connected apps, produce files) and adds a plan you approve first, recurring tasks and triggers from apps like Gmail and Slack.
This guide covers what ChatGPT agent did, where each of those jobs lives now, and a task brief you can paste into Work so the first run comes back right.
On this page
- What was ChatGPT agent?
- What replaced ChatGPT agent mode?
- How to use ChatGPT Work for agent-style tasks
- Where each old agent task goes now
- A task brief template for ChatGPT Work
- Worked example: a weekly competitor check
- Checklist before you hand over a task
- Questions people ask
What was ChatGPT agent?
OpenAI launched ChatGPT agent on 17 July 2025. It joined two earlier tools, Operator (which clicked through websites) and deep research (which read and summarized many sources), into one mode that could think, browse and act in the same task (OpenAI announcement). You switched it on from the tools menu or by typing /agent in the message box, described the task, and it worked in its own virtual browser.
A few things made it different from a normal chat:
- It could take actions as well as answer: fill in forms, compare prices, build a slideshow or a spreadsheet, and hand back the file.
- When a site needed a login, it paused and let you take control of the browser. Screenshots were not captured while you typed, to protect passwords.
- It asked before high-impact actions such as purchases, and on some sites it needed you to keep watching while it worked (OpenAI Help Center).
- At launch, Pro users got 400 agent messages a month and other paid plans got 40.
That article on the help center now opens with one line: ChatGPT agent is no longer available. It gives no retirement date, and older blog posts and videos about agent mode still rank, which is why so many people search for a feature they can no longer find.
What replaced ChatGPT agent mode?
ChatGPT Work replaced it. OpenAI introduced Work on 9 July 2026 as "an agent for longer, more involved tasks" (ChatGPT release notes). The help center describes it as an agent for longer, multi-step work and finished deliverables, while normal Chat stays for fast questions (ChatGPT Work and Codex).
Here is how the two compare on the points that matter day to day, based on OpenAI's own pages:
| Question | ChatGPT agent (retired) | ChatGPT Work (now) |
|---|---|---|
| How you start it | Tools menu or /agent in chat | Switch to Work at the top of the desktop app, or pick Work from the menu on web and mobile |
| Before it starts | Started straight away | Plan mode: it gathers context, asks questions and writes a step-by-step plan you approve or change |
| What it hands back | Reports, slideshows, spreadsheets | Documents, spreadsheets, presentations, reports and interactive sites |
| Your apps and files | Connected apps as data sources | Connected apps and plugins; on desktop, a local folder you choose |
| Recurring work | Scheduled agent runs counted against the monthly cap | Scheduled tasks, plus tasks triggered by new Gmail messages, Slack messages or GitHub pull requests |
| Approvals | Confirmation before high-impact actions | Approve or decline actions with on-screen controls (spoken approval is not supported) |
One change catches people out. Agent mode lived inside a normal chat. Work is a separate view, so if you look for agent in the tools menu of an ordinary conversation, you will not find it. Switch to Work first.
How to use ChatGPT Work for agent-style tasks
If you used ChatGPT agent for research, forms or files, these are the steps in Work, as the help center lists them:
On the desktop app
- Open the app and choose ChatGPT in the top-left menu.
- Toggle to Work at the top of the page.
- Start a new chat or open an existing project, then add files and describe the task.
On the web or mobile
- Open ChatGPT and choose Work from the menu.
- Start a new chat or open an existing project.
- Add files and context, then describe your task.
On desktop, Work can open a local folder and only get access to the files a task needs. On the web and mobile it cannot reach files on your computer, so upload what it needs or connect the app where the files live.
Which plans get ChatGPT Work?
The release notes say Work went to paid plans except Free and Go. The ChatGPT Work page says it is available to all plans on desktop, and to Plus, Pro, Business, Enterprise and Edu on the web and mobile. On company plans, Work is on by default unless an owner or admin turns it off. Usage works like Codex: how much you get depends on the task, the size of inputs and outputs, and your settings.
Where each old agent task goes now
ChatGPT agent did several kinds of jobs. They did not all move to the same place. Use this map:
| What you used agent mode for | Where to do it now |
|---|---|
| Research across many sites, ending in a report or deck | Work, with plan mode on so you can fix the scope first |
| Filling in forms or clicking through a website | Work's browser, approving each action on screen |
| A check that repeats every day or week | A scheduled task, managed from the Scheduled page |
| Acting when an email or Slack message arrives | An event-triggered task in Work (Gmail, Slack or GitHub) |
| The same job run by a whole team | A workspace agent on Business or Enterprise |
| Writing or fixing code | Codex, which stays separate for software work |
| A quick answer or a one-off draft | Normal chat. It is faster and uses less of your allowance |
ChatGPT scheduled tasks: limits to know
Scheduled tasks now have their own limits on how many can be active at once: 3 on Free and Go, 5 on Plus, 10 on Business and Edu, and 15 on Pro and Enterprise. Free users can run a task at most once a day; paid plans can run hourly. Event-triggered tasks from Gmail, Slack or GitHub need Plus, Pro, Business, Enterprise or Edu (Scheduled tasks in ChatGPT).
ChatGPT workspace agents for teams
If your team relied on one person's agent runs, look at workspace agents. On eligible Enterprise and Business workspaces you can build an agent with its own instructions, apps and files, test it, share it, run it on a schedule or start it through an API. Actions that change data in an app default to "Always ask" (workspace agents). Before you roll one out, agree the ground rules: our free AI policy template covers what staff may connect and approve.
A task brief template for ChatGPT Work
Agent runs usually go wrong for a boring reason: a one-line request with no finish line. Plan mode helps, because Work asks questions before it starts. You still save a round of questions by giving it the answers up front. Copy this and fill in the brackets:
Goal: [one sentence: what should exist when you finish]
Who it is for: [reader and what they will do with it]
Inputs: [files, folders or connected apps to use; anything to ignore]
Sources: [sites you trust; sites to avoid]
Steps I expect: [2 to 5 steps, or "propose a plan first"]
Output: [format, length, file type, file name]
Done means: [3 checks the result must pass]
Ask me before: [sending, buying, deleting, posting, sharing]
Do not: [log in to X, contact anyone, change existing files]
If you get stuck: [stop and tell me what is missing]The "Ask me before" and "Do not" lines matter most. They turn Work's approval step from a surprise into something you planned. If you want a second opinion on the brief itself, paste it into Roast My Prompt and fix what it flags.
Worked example: a weekly competitor check
This is an example set-up, not a result. Say you used agent mode every Monday to check three competitors' pricing pages and paste the changes into a sheet. In Work it becomes a scheduled task:
- Write the brief. Goal: a table of plan names and prices for three named competitors, with a column saying what changed since last week. Sources: only their pricing pages. Done means: every price has the page link next to it.
- Run it once by hand in Work and approve the plan. Check every row against the live pages.
- Turn it into a scheduled task for Monday mornings, with push or email notifications on in Settings > Notifications.
- Check the count. One weekly task uses 1 of your active task slots. On Plus that is 1 of 5, so 4 are left. A weekly task runs about 52 times a year (52 weeks), which is roughly 4 or 5 runs a month.
If a check only needs to happen when something arrives, such as a supplier's price-change email, use an event-triggered task on Gmail instead of a fixed schedule. It runs when the email lands, not every Monday whether anything changed or not.
Not sure which of your weekly jobs are worth handing over at all? The free AI task delegator sorts a list of tasks into what to delegate and what to keep, and the AI delegation scope tool helps you write the limits for each one.
Checklist before you hand a task to ChatGPT Work
- The brief says what "done" looks like, in checks you can tick.
- Work only has the files and apps this task needs. On desktop, grant one folder, not your whole drive.
- Anything that sends, buys, deletes or posts is on the "ask me before" list.
- You read the plan before you approve it, and you change it if a step looks wrong.
- You know the output will be checked by a person before it goes to a client or a boss.
- Your company allows the apps you are about to connect. If there is no policy yet, ask.
- For recurring tasks, you know where to pause them: the Scheduled page.
Web pages can contain hidden instructions aimed at AI tools. OpenAI built prompt injection monitoring into agent mode for this reason, and it is still worth keeping Work away from logins and payments it does not need.
What still needs you
Agent mode could not visit every website, and it paused for logins. Work follows the same pattern of asking for approval on actions. That means some tasks will never run fully unattended, and that is fine. Plan for three human moments: approving the plan, approving any action that changes something outside ChatGPT, and checking the result. If a task needs more than that from you every time, it may be quicker to do it yourself or to split it into smaller tasks.
Want Work set up for your actual job?
Custom AI for Your Job gives ChatGPT, Claude, Gemini or Grok one ready-made helper for each recurring task in your week, with briefs and limits written for your role.
Questions people ask
Is ChatGPT agent still available?
No. OpenAI's help center says ChatGPT agent is no longer available and points people to ChatGPT Work for longer, multi-step tasks and finished deliverables. Work can browse, use connected apps and files, and produce documents, spreadsheets and presentations. It also writes a plan you approve before it starts, which agent mode did not do.
What is ChatGPT Work?
ChatGPT Work is an agent inside ChatGPT for longer, more involved tasks. OpenAI introduced it on 9 July 2026. You give it a goal, it gathers context, asks questions and proposes a step-by-step plan, then does the work across your files and connected apps. It can hand back documents, spreadsheets, presentations, reports or interactive sites.
How do I use ChatGPT agent mode now?
Typing /agent or picking agent from the tools menu no longer works, because agent mode has been retired. Switch to Work instead: in the desktop app, toggle to Work at the top of the page; on the web or mobile, choose Work from the menu. Then add your files, describe the task and approve the plan it suggests.
Can ChatGPT Work run tasks on a schedule?
Yes. Scheduled tasks let ChatGPT handle recurring work, and on Plus, Pro, Business, Enterprise and Edu plans a task in Work can also start when something happens in Gmail, Slack or GitHub. Each plan has a cap on active tasks, from 3 on Free and Go to 15 on Pro and Enterprise. You manage them from the Scheduled page.
What happened to ChatGPT Operator?
Operator was folded into ChatGPT agent when agent launched in July 2025, so its website-clicking skills became part of agent mode. With agent mode now retired, that browsing and form-filling work happens in ChatGPT Work, which asks you to approve actions on screen before it changes anything outside ChatGPT.
Which plans include ChatGPT Work?
OpenAI's Work page says it is available to all plans in the desktop app, and to Plus, Pro, Business, Enterprise and Edu on the web and mobile. On company plans it is switched on by default unless an owner or admin turns it off. Usage depends on the task and follows the same structure as Codex.
Leave a comment: