ChatGPT agent: what it actually does

The ChatGPT agent can combine research, file analysis, and actions on supported websites to produce a concrete result, such as a spreadsheet or report. As of October 24, 2026, however, the former ChatGPT agent is no longer available: OpenAI directs users to ChatGPT Work, which handles longer tasks and deliverables. What you can delegate still depends on the available tools, the permissions you grant, and your ability to check the results.

ChatGPT agent has changed: look for Work now

The official ChatGPT agent help page announces that it is no longer available and recommends Work. Yet the page still includes sections describing the former experience, its plans, and its quotas. Because both appear together, take care not to read those older details as current subscription terms.

When it launched on July 17, 2025, ChatGPT agent brought browsing, research, and code execution together in a virtual computer. Its capabilities included preparing presentations and spreadsheets, according to OpenAI’s original announcement. That history explains the demonstrations and tutorials still circulating, but does not guarantee that the same features remain available in the same form today.

The useful distinction is between receiving an answer and having a task carried out. An AI agent selects intermediate actions to reach a goal, so you need to examine what it actually accessed, changed, or produced. Our article on the difference between an agent and a chatbot explains this shift in how the software works.

This guide covers the capabilities currently described in the documentation and offers sample requests. It does not report personal testing or measured success rates.

Tasks you can delegate in practice

OpenAI presents Work as an experience for researching, analyzing, and creating documents, spreadsheets, presentations, or reports. It can draw on files and connected apps, depending on the access available, as described in the ChatGPT Work documentation. The practical value is in requesting a deliverable you can check, rather than a series of suggestions you must carry out yourself.

Here are sample requests to adapt to your needs. They do not imply that every task will succeed on every account or website:

Need	What to request	What to verify
Support a purchasing decision	Gather offers that meet explicit criteria	Prices, availability, fees, and exclusions
Process documents	Extract deadlines from a set of documents and create a table	Whether each entry matches the source passages
Prepare for a meeting	Organize the supplied material into a structured briefing	Names, dates, and missing information
Prepare an online submission	Research the options and prepare the authorized form	The website, entered fields, and final commitment

For product or supplier research, for example, ask it to separate confirmed specifications from information it could not find. For document analysis, require a reference to the file and passage used. These requirements make mistakes easier to spot during review.

The right scope also depends on how much work checking the result will take. If verification means doing the entire task again, delegation offers little benefit. A clearly bounded assignment, with accessible sources and a specific output format, is easier to evaluate than an open-ended goal such as managing your business.

The browser works in a separate environment

Work’s cloud browser runs on a remote computer. It can visit pages, click, and fill out forms on supported websites, and a task can continue after you shut down your device. It does not automatically inherit the logins from your personal browser.

The cloud browser documentation also distinguishes permission to access a website from confirmation of an action that creates a commitment. Allowing access to a domain therefore does not authorize every payment or reservation. Some websites block automated access, and manual intervention may still be necessary.

For supported logins, a secure form passes your credentials to the remote browser without showing them to the model. Sessions can persist, so review retained browsing data after a sensitive task.

These distinctions have practical consequences. Being signed in to a service in your usual browser does not mean Work has the same access. If a form appears complete, ask whether the information has merely been prepared, saved as a draft, or submitted.

For more on this category of tools, see our guide to agents that use a computer. How the agent interacts with the site matters as much as the quality of its written output: a good summary does not establish that an external action succeeded.

A method for getting a result you can check

Before you begin, confirm that Work appears in your account, that the necessary files are available, and that you have permission to share them. Open the source documents as well, so you can check the deliverable against them.

1. Define the expected output. Specify the format, scope, and intended audience: a comparison table, an internal briefing, or a draft form.
2. Set the source boundaries. Identify the files to use, the acceptable websites, and any information to exclude.
3. Specify the permitted actions. Distinguish reading, preparation, editing, and submission. For an initial trial, limit the assignment to a deliverable you can review.
4. State the success criteria. Ask for references, unknowns, and a list of unfinished steps.
5. Review before proceeding. Examine the result, then authorize only the next necessary step.

Here is a sample instruction to adapt to your documents:

Using the attached quotes, prepare a table comparing the services,
exclusions, timelines, and payment terms.
For each item, identify the source document.
Write "not specified" when information is missing.
Do not contact any supplier or accept any offer.
Finish with the points that need my review.

This wording establishes the scope of the assignment. It also helps you spot an expansion of that scope: contacting a provider to fill a gap in a table is a new action with different consequences.

To assess whether the tool is useful, also track the time you spend on corrections and approvals. The displayed execution time alone does not tell you how much work you saved.

Common problems: access, errors, and budget

Work is missing, or a tool is unavailable. Check your plan, the rollout, and your workspace permissions. The documentation describes different kinds of access; having the standard chat interface does not guarantee access to every agent feature. Avoid repeatedly signing out and back in without first identifying what is missing.

The task seems stuck. Look for a request for clarification, a login, or approval. If a website rejects automation, complete that part manually or supply the authorized data in another format. A partial response should explain what remains inaccessible.

The deliverable looks consistent but contains errors. Check the decisive facts against the original documents. A hallucination might appear as a contract term that was never stated, an inferred date, or an explanation expressed with too much certainty. Ask the tool to flag unknowns instead of filling in the gaps.

Your available usage is dropping quickly. Current usage guidance does not translate into a fixed number of tasks: the model, task complexity, and settings affect consumption. Check your usage settings and account terms, following the Work usage limits documentation.

When budgeting, separate the subscription, any additional usage charges, and the human time spent supervising the work. Do not treat ChatGPT’s entry price as the price of an unlimited agent service. Our article on the real cost of an AI agent outlines the expenses to track.

Permissions limit the consequences of an error

A connected agent may encounter private information and external content within the same assignment. A prompt injection attempts to turn that external content into an instruction: a page or document tries to push the agent beyond its original assignment.

OpenAI describes this risk and recommends measures including reducing access, making requests specific, and reviewing actions before confirming them in its overview of prompt injections. These precautions reduce exposure without guaranteeing that errors will never occur.

In practice, start with public data or redacted copies. Grant access to a sensitive app only when the task requires it. Before sending anything, check the recipients and attachments; before making a reservation, check the dates, amount, and applicable terms.

Approval needs to cover the exact action. Permission to research availability does not establish a budget, cancellation terms, or the identity of the person making the commitment. The guide to securing an AI agent can help you formalize these boundaries.

If your priority is a repeatable process, detailed permissions, or control over the installation, also consult the AI agent directory. To evaluate another assistant, use the dedicated pages, including the ChatGPT and Claude comparison.

Checklist before delegating

ChatGPT agent now refers to a former experience. To carry out tasks today, check which Work capabilities are available in your account. Start with an assignment whose success you can recognize, then expand its scope after reviewing the result.

• The deliverable and acceptance criteria are explicit.
• Accessible sources and accounts are limited to what the task needs.
• Permitted actions and required approvals are defined.
• Usage costs and review time are tracked.
• The result distinguishes verified information, unknowns, and unfinished actions.

Frequently asked questions

What happened to ChatGPT's agent mode?

The official help page now says ChatGPT agent is no longer available and directs users to ChatGPT Work. Older tutorials showing an agent menu or a dedicated command may therefore describe an outdated interface.

Is ChatGPT agent free?

Check access to ChatGPT Work and its cloud browser rather than looking for a subscription to the former agent mode. The cloud browser documentation describes access through eligible paid plans, excluding Free and Go, with availability varying by rollout and permissions.

Should I give ChatGPT my password?

Do not paste it into the conversation. For supported sites, the cloud browser provides a secure sign-in form; check the site's address and the information it requests before continuing.

Can ChatGPT keep working when my computer is off?

A task using the cloud browser can continue on ChatGPT's remote computer after you shut down your device. It may still pause to request a login, clarification, or confirmation; this behavior does not apply to every kind of local task.