· AI Seedling · Beginner Guides · 11 min read
ChatGPT Agent Mode: What It Was and What Replaced It
Agent mode is no longer available in ChatGPT. What it did, why it existed, and how to use Work, the feature that replaced it.
If you searched for agent mode because someone mentioned it, the first thing you need to know is that it is no longer available.
ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables.
Source: OpenAI, ChatGPT agent, checked October 27, 2026. The page had been updated four days before that check.
So the honest answer to “what is agent mode” has two parts. It was a mode where ChatGPT operated a browser and completed tasks on your behalf, and it has been replaced by a feature called Work. This page explains what agent mode did, why it mattered, and how to do the same thing today.
What agent mode actually did
Ordinary ChatGPT answers. Agent mode acted.
You gave it a goal, and it opened a virtual browser, clicked through websites, filled in forms, read pages, and worked through the steps itself. It paused to ask you questions or get confirmation when a step looked consequential. It could run for minutes rather than seconds.
The difference is easier to see with an example. Asking “find me a restaurant for six people on Friday” in ordinary chat gets you a list of suggestions. Agent mode could check your calendar, look at recent emails for who is coming, search a booking site, and hold a table — pausing before anything that committed you to something.
That capability came with a specific shape: a virtual browser the agent controlled, screenshots it took to “see” pages, and a requirement that you supervise anything sensitive.
How it differed from ordinary chat
| Ordinary chat | Agent mode | |
|---|---|---|
| What it produces | Text you read and act on | Actions taken on your behalf |
| Who does the clicking | You | The agent, in a virtual browser |
| Typical duration | Seconds | Minutes, sometimes longer |
| Failure mode | A wrong answer | A wrong action, or an action on the wrong account |
| Your role | Judge the answer | Supervise the process |
That last row is the important one. A wrong answer costs you the time to notice it. A wrong action can send an email, change a setting, or book something before you have read the confirmation.
What replaced it: Work
OpenAI now describes ChatGPT as having two main experiences, plus a separate view for coding:
Chat is for fast, conversational assistance and everyday questions. Work is an agent designed for longer, multi-step work and finished deliverables. Codex remains dedicated to software development and technical work.
Source: OpenAI, ChatGPT Work, checked October 27, 2026.
If you only remember one line from this page, make it this one. Chat answers questions. Work finishes jobs. Codex writes software.
Work carries the agent behaviour forward rather than dropping it:
Work can also run once, repeat on a schedule or trigger, or monitor for changes through Scheduled Tasks. You can review progress, answer questions, change direction, and approve important actions.
Source: same page, checked October 27, 2026.
That last clause matters more than it sounds. “Approve important actions” means the design assumes you are still in the loop for anything consequential. Work is not a hands-off autopilot, and treating it as one is where people get hurt.
Why the names keep changing
If you have read about Operator, agent mode, and Work in the same month, you are not confused. The product has been renamed more than once.
Operator came first: a separate product that operated a browser. Its functionality was later folded into agent mode inside ChatGPT, and the Operator website was closed. Agent mode then gave way to Work, which now sits alongside Chat as one of two main experiences.
The pattern matters because of what it does to tutorials. A guide written six months ago may describe a menu that no longer exists, a product name that has been retired, or a limit that has since changed. When instructions do not match what you see on screen, the instructions are usually the outdated part.
The practical rule: trust the interface in front of you and the current help documentation, not a blog post with a screenshot from another version.
What Work can actually produce
Research a topic, analyze information, or create a document, spreadsheet, presentation, report, or Site.
Source: same page, checked October 27, 2026.
In plain terms, Work is aimed at things that take a while and end in a file:
| You want | Work is a reasonable fit | Why |
|---|---|---|
| A research summary with sources | Yes | It can read several pages and compile them |
| A spreadsheet from messy notes | Yes | Structured output is what it is built for |
| A slide deck from a document | Yes | Multi-step, ends in a deliverable |
| A quick definition | No | Chat answers this in one message |
| A decision about hiring someone | No | No amount of task execution replaces judgment |
| Anything with your bank login | No | The risk is not worth the convenience |
How to start a Work task
The general shape is the same as agent mode was:
- Open ChatGPT and look for the mode selector, or switch to Work where your interface offers it.
- Describe the task and the finished deliverable you want — “a two-page summary with sources”, not “look into this”.
- Answer its clarifying questions. This is the step people skip, and it is the step that determines whether the output is usable.
- Watch the progress view. You can change direction mid-task.
- Approve or decline anything it flags before it completes.
- Read the result as if a colleague produced it and you have to sign it.
Availability depends on your plan and workspace. Work is offered on paid plans; free accounts do not have it. If you cannot find the mode selector, that is usually the reason rather than a broken interface.
A worked example you can try without logging in
Here is what a Work task looks like in practice, using material that requires no accounts.
The prompt:
I have notes from three meetings below. Produce a one-page brief with four sections: decisions made, open questions, who owns each next step, and dates mentioned. Do not invent an owner or a date that is not in the notes. If a section has nothing, say so.
What happens:
- Work reads the notes and may ask which meeting is the most recent if the dates are ambiguous.
- It produces a draft with the four sections.
- You read it and find an action item assigned to a name that appears with two different spellings.
- You correct it and ask for one revision.
Why this task suits Work rather than Chat: the output is a document with a fixed structure, the source material is longer than one message, and every claim can be checked against the notes. That combination is the sweet spot — enough steps to justify the agent, enough verifiability that you can catch a mistake.
Why it needs no logins: everything is in the prompt. The moment a task requires an account, you are accepting a new category of risk, and that should be a deliberate decision rather than a side effect of convenience.
Usage limits worth knowing
Agent-style work consumes more than a chat message, and the limits are set accordingly. In agent mode the documented monthly caps were 40 messages on Plus, 400 on Pro, and 40 on Business and Enterprise. Work follows a different structure tied to your plan.
Two details from the agent mode documentation are worth carrying over, because they apply to any agent product:
- Only the requests you initiate count toward the limit. Clarifying questions and authentication steps in the middle of a task are not charged.
- Every separate invocation counts, including ones started by a scheduled task. A daily task quietly uses thirty of your monthly allowance.
If Work is not available on your plan
Free accounts do not have Work, and some workspaces disable agent features for everyone. You are not stuck — you do the steps yourself with Chat as an assistant.
The workflow that keeps most of the benefit:
- Ask Chat to help you plan the task as a numbered list.
- Do the steps yourself, pasting results back as you go.
- Ask Chat to assemble the pieces into the final format.
You lose the automation and keep the structure. For a one-off task that is often the better trade anyway, because you see every step and notice when one goes wrong.
The risk that matters: prompt injection
This is the part most beginner guides skip, and it is the reason agent products need supervision.
When you sign ChatGPT agent into websites or enable apps, it can access sensitive data like emails, files, or account settings, and perform actions on your behalf (e.g., sharing files, modifying account settings). This creates potential privacy risks, including “prompt injection” attacks.
Source: OpenAI, ChatGPT agent, checked October 27, 2026.
OpenAI’s own example is worth repeating because it is so ordinary:
For example, you might ask the agent to find a restaurant for a group dinner by checking your calendar and recent emails. While researching, it might encounter a malicious comment — essentially a harmful piece of content designed to trick the agent into performing unintended actions — directing it to retrieve a password reset code from Gmail and send it to a malicious website.
Source: same page, checked October 27, 2026.
Read that again. The attack does not come from you. It comes from a web page the agent reads while doing something else. A comment on a restaurant review page is enough.
The safeguards exist — confirmation prompts, refusal patterns, injection monitoring, and a watch mode on some sites — but OpenAI states plainly that these measures “don’t eliminate all risks.”
What you can do, in order of how much difference it makes:
- Log in to as few things as possible. An agent cannot leak access to an account it was never given.
- Turn off apps you are not using for this task. Every connected app is another door.
- Never type a password into the chat. Use the browser takeover option when a login is needed, because screenshots are not captured while you are in control.
- Avoid vague prompts. “Check my email and handle everything” is an instruction to act on whatever it finds.
- Stop the task when something looks wrong. You are allowed to interrupt.
When to use Chat, and when to use Work
A simple test: if the answer fits in a message, use Chat. If the output is a file, use Work. If the task requires clicking through sites you would not hand to a stranger, use neither.
| Situation | Use |
|---|---|
| Explain a concept | Chat |
| Draft an email | Chat |
| Compile research from five pages into a summary document | Work |
| Build a spreadsheet from a pile of notes | Work |
| Book something that requires your payment details | Neither — do it yourself |
The third column is not a limitation of the product. It is a judgement about what you can check afterwards.
Questions beginners often ask
Can I still use agent mode?
No. It has been replaced. If a tutorial tells you to type /agent, that tutorial is out of date. Look for Work in your mode selector instead.
Is Work free?
No. Work is offered on eligible paid plans. Free accounts have Chat only.
Does Work give me sources?
Yes. Agent mode documented that outputs include source links or screenshots, and that behaviour carries into the Work product. Check them anyway — a citation you have not opened is not evidence.
What happens to my screenshots and browsing history?
They stay in your conversation history until you delete them, and deleting the chat deletes the associated screenshots. On business plans, your data is not used for training by default. On Plus and Pro, check your data controls setting.
Is it safe to let it log into my accounts?
That depends entirely on which accounts. A throwaway account for a shopping site is a different risk from your primary email. Treat every login you grant as a door you have opened, and close the ones you do not need.
When to ask a person
Ask a person before letting an agent act on anything involving money, health, legal matters, employment, or another person’s information. Ask before granting access to a shared workspace account. Ask when a task’s outcome would be difficult to undo.
An agent can save you an afternoon of clicking. It cannot take responsibility for what it clicked on.
Keep learning in small steps
Start with a task whose result you can check in a minute: a short summary of three pages you have already read. Compare what Work produced against what you know. If it missed something obvious, that tells you more about its limits than a successful demo does.
Then try one task with a real deliverable, still low-stakes. A meeting agenda, a one-page comparison, a cleaned-up list. Keep the prompts that worked and delete the ones that produced confident nonsense.
The habit that matters is the same one that matters for ordinary chat, only more so: you are responsible for the output, and the tool’s fluency is not evidence that it is right. If you want a structured way to check where you stand, the AI Level Test takes three minutes and covers exactly this kind of judgement.
If Work itself stops responding, or you get an error you do not recognise, the ChatGPT troubleshooting checklist covers each message and what to collect before contacting support.
