ChatGPT agent mode no longer exists as a separate feature. OpenAI’s own help page about agent mode now opens with the notice that ChatGPT agent is no longer available and that you should use ChatGPT Work for longer tasks with several steps. If you type /agent in the message box, nothing happens any more.
The capability itself has not disappeared but moved. ChatGPT Work, launched on 9 July 2026, does what agent mode did and more, but in a different place and with different limits. The same pattern played out earlier with ChatGPT canvas, which was absorbed into writing blocks within the chat in May 2026.
If you used agent mode, more has changed than a name. Scheduled tasks that relied on the agent stop running, the usage limits work differently and the part that did the browsing now sits in a separate environment. This article sets out exactly what has changed, what agent mode could do, and when an AI agent is the right choice for a task at all.
Does ChatGPT agent mode still exist?
ChatGPT agent mode has been discontinued and replaced by ChatGPT Work, which OpenAI launched on 9 July 2026 on the GPT-5.6 models. Since then, OpenAI’s help page about agent mode has started with the statement that the feature is no longer available, pointing you to Work for longer tasks and to the cloud browser for browser work.
Confusingly, the rest of that help page still carries the old instructions. It still describes how to select agent mode from the Tools menu, what the monthly limits are and how watch mode works, while only the first line says the whole thing is gone. If you scan the page instead of reading it, you will conclude the feature is still running, which explains why many guides online still describe the old situation.
The discontinuation also came without a separate announcement. Agent mode appeared in July 2025, absorbed Operator and Deep Research along the way, and disappeared a year later in a product launch in which the words agent mode barely featured. Users mainly noticed because the /agent command suddenly stopped doing anything.
What is ChatGPT Work?
ChatGPT Work is the successor to agent mode; it started on GPT-5.6 and has Codex built in. GPT-6 Astra has been the top model in Work since 3 September 2026, for Plus, Pro and Business, and GPT-6 Sol and Luna have run in it since 22 September 2026. GPT-6.1 Sol has been rolling out since 29 September 2026 for Plus, Pro, Business and Enterprise, and GPT-6.1 Terra has been rolling out since 26 October 2026. In Enterprise and Edu, an admin switches those models on. Work carries out tasks with several steps and delivers a finished result instead of an answer in the chat. Work can be started and managed on web, mobile and desktop, so a task is no longer tied to one device.
Its reach is wider than that of agent mode. Where the agent was limited to a virtual browser, Work can also use local files and applications, or a built-in browser to access websites and online files. Plugins connect systems and apps to Work, and with a separate beta you can build interactive reports and dashboards with it.
At launch, Work rolled out to the paid plans, except Go. Free was excluded at the time too, but OpenAI’s pricing page now lists limited access to Work in the desktop app for Free.
The desktop apps changed too. OpenAI merged the Codex app and the ChatGPT app into one new desktop application for Mac and Windows, with the old app continuing under the name ChatGPT Classic. If you already had Codex installed, you can update it to the new app rather than reinstalling.
At Anthropic, the equivalent is Claude Cowork, which takes a similar step by handling tasks on its own; since 6 October 2026 it has run in the cloud by default on Pro, Max and Ultra instead of on your own computer. Since 16 September 2026, what Cowork does has been available from any conversation in Claude, with a gradual rollout to Pro and Max. Both companies are moving towards the same model, in which you hand off a task and review the result afterwards.
What exactly did agent mode do?
Agent mode carried out tasks on its own in a temporary workspace, instead of giving an answer in the chat. You described in plain language what needed to happen, after which ChatGPT determined the steps itself and worked through them. Three components shaped how that felt in practice.
The virtual browser
The virtual browser was the heart of agent mode. In a sandboxed environment, ChatGPT visited websites, clicked through pages, filled in forms and collected data, using screenshots of that window to determine where it was and what the next step should be.
A task therefore typically took five to thirty minutes, depending on its complexity. You could watch while it ran, steer it or stop the task. The session did have to stay active, which meant a long task tied you to the screen, or at least to an open window.
Built-in limits and watch mode
The built-in limits meant the agent refused certain actions outright. Making payments and entering passwords were excluded, and for sensitive websites the agent switched to watch mode, in which you approved every action before it continued.
In setting those limits, OpenAI explicitly mentioned the risk of prompt injection, in which a malicious instruction on a web page tries to get the agent to do what the attacker wants instead of what you asked. That risk does not go away with Work, because an agent with access to your email and files is by definition exposed to it.
Takeover for logins and payments
Takeover let you temporarily take control of the virtual browser. If the agent reached a login screen or a payment step, it paused and you took over, with no screenshots recorded while you were at the controls.
You then handed control back and the agent picked up where it had left off, although that did not always work and a task sometimes had to start again. Cookies were kept between sessions, just as in an ordinary browser, which was convenient but also meant login details sat in a place you had to clean up yourself.
What changes if you used agent mode?
If you used agent mode every day, you will notice the transition in three ways. The /agent command does nothing any more, the button has disappeared from the Tools menu, and scheduled tasks that called an agent no longer run as they did.
That last one is the most annoying. Recurring tasks that relied on agent mode have to be set up again in Work or in Codex, and there is no automatic migration. That is work you have to do yourself, and it pays to check first which of your scheduled tasks are still running at all.
The usage limits are also different. Agent mode had a monthly budget per plan in which only your own starting prompts counted and follow-up questions were free. If you now work with long-running tasks, check how Work counts them before you set up a recurring process that gets stuck halfway through the month.
Checking whether you have Work takes a moment. Look in the menu next to the message box to see whether Work is listed as an option. If not, your plan is not yet included or the rollout has not reached you yet. There is little point in looking for agent mode in the meantime, because that button is not coming back.
When do you hand a task to an AI agent?
Handing a task to an AI agent works as long as the assignment is well defined and the result can be checked. Comparing twenty suppliers and putting the outcome in a table is exactly the kind of work an agent is made for. Keeping up with your email and deciding for yourself what gets answered is not.
The principle behind this comes from banking practice: not every task has the same risk profile. An agent that collects and summarises public information can proceed with a light check. An agent with access to client files, mailboxes or an accounting system deserves a heavier one. You make that assessment in advance per type of task, not afterwards per incident.
Three questions for every AI task make this practical: what data goes in, what claims do you base on the outcome and who is responsible for the result? With an agent, the first question matters even more than in an ordinary chat, because the agent decides for itself which sources it opens along the way. Which information you do and do not give a tool is part of responsible AI use and should be recorded before anyone lets an agent loose on a client file.
The check on the final result still stands. An agent working on its own does not give an error message if a website behaves unexpectedly; it delivers a result in which a step is missing. An AI hallucination is even harder to spot in agent work, because the output looks like the product of real research. The four-eyes principle from my years at major banks applies here in full: nothing with real stakes goes out without a second check.
For knowledge-intensive firms, that means a short list of suitable tasks and a longer one of unsuitable tasks. Market scans, supplier comparisons, collecting public company data and summarising several public sources lend themselves well, because the source can be checked and the error is visible. Drafting a deed, writing an advice letter or assessing a client file does not, because there the error only shows when someone with expertise reads it, and that person could just as well have done it themselves.
If a task recurs often enough, an agent is not always the right tool either. The move from one-off prompts to fixed workflows in my own work came from exactly that question: what is the same every week is better captured in a script or an API integration than handed off each time to an agent that tackles it slightly differently. You learn to build such a fixed integration yourself in the LearnLLM AI Automation Course.
Frequently asked questions about the successor to ChatGPT agent mode
When you move from ChatGPT agent mode to Work, you will find that the practical rules differ on a few points.
How does ChatGPT Work count towards your usage limit?
ChatGPT Work counts towards your usage limit in the same way as Codex, and Work and Codex both draw on the allowance that comes with your plan. According to OpenAI, usage depends on the size of the task and the reasoning level, and longer conversations and tasks count towards it as usual. On Plus, use of GPT-6 Astra is limited, with the option to buy extra credits afterwards.
Can ChatGPT Work access files on your computer?
ChatGPT Work can only access files on your computer in the desktop app, and then only with your permission and if your plan and workspace allow it. According to OpenAI, Work has no direct access to local files on web and mobile. OpenAI advises giving access only to the files a task needs, and states that messages and task context may be stored in the cloud, even if the work runs locally.
Can ChatGPT Work start automatically when a new email arrives?
Yes, ChatGPT Work can start automatically when a new email arrives, through a task that responds to an event, such as a new message in Gmail. According to OpenAI, this works the same way for new messages in a Slack channel and for activity around pull requests in GitHub. Such tasks are not available on Free and Go, and in Enterprise and Edu an admin has to switch them on first.
What does this mean for the way you work?
Agent mode existed for less than a year, and that is the real signal in this story. A feature presented in July 2025 as a leap forward had been absorbed into something else by July 2026. In the same period, canvas disappeared from the default models and OpenAI discontinued the Atlas browser on 9 August 2026; the browser features have been in the desktop app since then.
I briefly tested ChatGPT 5.5 after its release to place the model in the wider market, and the picture that emerged then applies again here. The changes are mainly in how the product is organised, less in what the model can do at its core. For the way you work, that means: expect things to move, not to stand still.
In practice, it comes down to two habits. Record which step you take and why, not which button you use, so that when a feature moves, you only need to find out where it now sits. And put your checkpoint on the result rather than on the tool, because that point does not change with a product update.
Firms that want to set this up in a structured way, with agreed rules per type of task and checkpoints that stay in place when a supplier changes something, will recognise this as the core of an AI implementation for businesses. If you first want to learn to work better with the tool yourself, including prompt structure and checking output, you will find that in the LearnLLM ChatGPT Course.
