How to Use ChatGPT Work (Formerly Agent Mode) for Real Tasks

ChatGPT agent mode no longer exists. OpenAI’s help center now states “ChatGPT agent is no longer available” and points users to ChatGPT Work for “longer, multi-step tasks and finished deliverables,” with a separate cloud browser for supported browser workflows. So if you searched for how to use ChatGPT agent mode, the practical answer is: open the Work tab next to Chat, describe the outcome you want, connect the apps it needs, approve its plan and important actions, and review the finished file.

This guide shows how to do that step by step, with real tasks that work well, prompts you can copy, what each plan includes, and the mistakes that waste your usage.

What happened to ChatGPT agent mode?

Agent mode launched in July 2025. It combined the earlier Operator browser agent with deep research, so ChatGPT could browse a virtual computer, fill forms and pause for your confirmation. In July 2026 OpenAI introduced ChatGPT Work, described in its release notes as “an agent for longer, more involved tasks.” A few weeks later, the ChatGPT agent help article gained a notice that agent mode is no longer available.

OpenAI also deprecated its Atlas browser, which stopped working on 9 August 2026, as it moved “browser-based agentic capabilities into ChatGPT and Codex.” The result is a simpler lineup:

What ChatGPT Work can do for you

Work is built for jobs that take many steps and end in something you can use. Instead of answering one question, it takes a goal, plans the steps, works through them and hands back a finished result. Based on OpenAI’s launch materials and press coverage, it can:

• Research a topic across the web and your connected sources, then write it up.
• Work across apps you connect through plugins, such as Slack, Google Drive, email or a CRM, by typing “@” and the app name.
• Create documents, spreadsheets, presentations and reports, and turn work into a shareable web page or app with ChatGPT Sites (in beta).
• Run tasks once, on a schedule, when an event happens, or to monitor for changes.
• On desktop, use local files and apps, a built-in browser, and Computer Use to click, type and move files on your behalf, with your permission.
• Ask you questions, show progress and request approval before important actions.

Think of it as a capable junior colleague rather than a magic button. The more clearly you define the outcome, the better the result. That is the same skill our prompt engineering guide teaches for normal chats.

What you need before you start

A plan that includes Work

Work rolled out first to Pro, Enterprise and Edu on web and mobile in July 2026, then to Plus and Business. Free users can use Work in the macOS and Windows desktop apps with limited usage. Plus and higher get GPT-6 Astra inside Work, according to OpenAI’s GPT-6 help article.

Full plan details are in our ChatGPT pricing guide, and our Plus vs Pro comparison helps if you are unsure whether to upgrade.

The desktop app (recommended)

OpenAI’s Academy quickstart for Work says to “download the desktop app and begin working on your first task.” The new desktop app for Mac and Windows combines Chat, Work and Codex. The previous app may still show up as “ChatGPT Classic,” so make sure you install the new one. Desktop is also where local files, the built-in browser and Computer Use are available.

Your apps and files ready

Decide which tools the task needs, such as Google Drive for source documents or your calendar for meeting prep, and connect only those. Plugins bundle skills and connected apps, and they keep each app’s own permission and “action approval” rules.

How to use ChatGPT Work step by step

1. Open the Work tab. In the desktop app or on chatgpt.com, select Work next to Chat. You will see a task box similar to the chat box, plus access to plugins and settings.
2. Describe one clear outcome. Say what the finished deliverable is, who it is for and what “done” looks like. For example: “Create a 10-slide competitor review of three Pune bakeries for my cafe launch, with prices, best sellers and a one-slide recommendation.”
3. Point it at the right sources. Type “@” to pull in a connected app such as Google Drive, or attach files. Tell it which sources to trust and which to ignore, for example “use only the attached sales sheet for numbers.”
4. Ask for a plan before it starts. Add “Show me your plan first and wait for my approval.” Work supports check-ins and approvals, and some guides describe a setting to always plan before executing. Edit the plan if a step is wrong or missing.
5. Set boundaries and approvals. State what it must not do: “Do not send any email or post anything. Draft only.” For browser tasks on sign-in sites, Work asks for confirmation before consequential actions, but say it explicitly anyway.
6. Let it work, and answer its questions. You can follow progress on desktop or mobile, change direction mid-task and answer clarifying questions. Long tasks can run for a long time, so you do not need to watch every step.
7. Review the deliverable carefully. Open the file, check every number against its source, and ask for fixes in plain language: “Slide 4 uses last year’s price list. Update it from the attached PDF.”
8. Schedule it if it repeats. Once a task works, turn it into a scheduled task, such as “every Monday at 9 am, summarize last week’s support tickets from the shared folder.” Run it once manually first.

Real tasks that work well (with prompts)

These are practical jobs for marketers, founders and freelancers. Copy the prompt, then adjust the details.

1. Competitor research brief

“Research the top five competitors for [product] in India. For each, list pricing, main features, target customer and one weakness. Put it in a spreadsheet and write a one-page summary with three opportunities for us. Cite sources.”

2. Weekly marketing report

“Using the attached export from our ad account and the analytics CSV, build a weekly report: spend, leads, cost per lead and top three campaigns, compared with last week. Output a 5-slide deck and a short email draft for my manager. Do not send anything.”

3. Meeting prep pack

“For my 3 pm call with [client], pull our last email thread from @Gmail and the proposal from @Google Drive. Give me a one-page brief: what they asked for, open issues and three questions to ask.”

4. Content calendar

“Plan a 4-week blog and Instagram calendar for a home bakery in Ahmedabad around Diwali. Include post ideas, captions and suggested posting days in a spreadsheet.” Our ChatGPT prompts for marketing collection has more ideas to feed into tasks like this.

5. Lead list clean-up

“Clean the attached lead sheet: remove duplicates, standardize phone numbers to +91 format, flag missing emails and add a column for city. Return the cleaned file and a summary of changes.”

If your workflow is a simple trigger and action, such as “when a form is filled, add a row and send a WhatsApp,” a dedicated automation tool is usually cheaper and more reliable. Compare options in our guides to automating tasks with Zapier and AI and the best AI automation tools.

Usage limits and costs: how Work is metered

Work does not use simple message counts. OpenAI measures Work and Codex usage in credits, and both draw from the same allowance. In the desktop app you can open Settings, then Usage and billing, to see your monthly credit usage, remaining allowance, reset timing and your top consuming chats, according to OpenAI’s Work and Codex usage help article. Each Work chat header also shows its lifetime credits.

OpenAI does not publish exact Work allowances per plan. What we know is that more complex tasks consume more, and that sub-agents, image generation and some background activity may not show in a chat’s total but still count. That means one sprawling “do everything” task can use far more than five focused ones. Our guide to ChatGPT limits by plan covers the chat side of your allowance.

Safety and privacy: rules to follow

An agent that can read your inbox and click buttons deserves caution. Follow these rules, especially on a business account.

• Connect the minimum. Give Work only the apps a task needs, and disconnect plugins you no longer use.
• Keep “draft only” as the default. Let it prepare emails, posts and record changes, then send or save them yourself.
• Approve consequential actions. Purchases, deletions, sending messages and changing CRM records should always need your approval.
• Watch for prompt injection. Web pages and documents can contain hidden instructions. If Work suddenly tries something you did not ask for, stop the task.
• Check workspace rules. Business and Enterprise admins can control access, connected tools and permitted actions. Write your own rules down using our AI usage policy template, and read our guide on AI data privacy for business before connecting client data.

Common mistakes with ChatGPT Work

1. Vague goals. “Help with marketing” produces generic output. Name the deliverable, audience and deadline.
2. Skipping the plan. Approving a plan takes a minute and prevents an hour of wrong work.
3. Trusting numbers blindly. Always check figures and sources before you share a report.
4. Using Work for quick questions. Chat with GPT-6 is faster and lighter on usage for simple answers.
5. Forgetting old schedules. Agent-era scheduled tasks need rebuilding in Work.
6. Relying on custom GPTs. OpenAI announced in September 2026 that custom GPTs will be retired with a migration path to plugins, so build new workflows on plugins and Work rather than on our older custom GPT guide alone.

Frequently asked questions

Next steps: start with one real task

ChatGPT Work is the successor to agent mode and is genuinely useful for research briefs, reports, decks and recurring admin, as long as you give it a clear outcome, ask for a plan and keep consequential actions under your control. Start with one task you do every week, run it manually, then schedule it once the output is reliable. Compare it with other options in our list of the best AI agents you can use in 2026. OpenAI changes plans, limits and feature names often, so confirm current access and pricing on the official ChatGPT site before you rely on it for client work.