Direct answer
ChatGPT agent mode was OpenAI's tool for delegated, multi-step tasks. OpenAI's current help center says the legacy ChatGPT agent is no longer available. Use ChatGPT Work for longer research, analysis, and finished deliverables. Use cloud browser inside Work for supported public-web actions, or switch to Codex for software development. This article explains the change and the current workflow as of 10 August 2026.
Key takeaways
- The product once called ChatGPT agent mode has been replaced by ChatGPT Work.
- Work is for longer, multi-step tasks and finished documents, spreadsheets, reports, presentations, or Sites.
- Chat remains the faster choice for questions, brainstorming, search, and ordinary conversation.
- Codex is the dedicated experience for repositories, code changes, tests, and developer tools.
- Cloud browser handles supported public websites but cannot sign in or complete payments at launch.
- Give every delegated task a clear outcome, inputs, constraints, approval points, and review criteria.
What happened to ChatGPT agent mode?
OpenAI introduced ChatGPT agent in July 2025 as a way to combine research, browsing, code execution, and online actions. Older articles therefore tell readers to open the plus or tools menu and select Agent mode. That instruction is now historical. The current OpenAI help page explicitly says the legacy ChatGPT agent is no longer available and points readers to Work for longer tasks.
What is ChatGPT agent mode?
ChatGPT agent mode was a supervised way to delegate an outcome instead of requesting one response at a time. A ChatGPT agent could plan steps, use tools, browse pages, work with data, and pause for clarification or approval. The useful idea survives in Work: you describe a deliverable, provide context and constraints, watch progress, answer questions, and review the result. It is not the same as giving an unsupervised system permanent authority.
Why searches also say ChatGPT agents
People use ChatGPT agents as a broad name for agent-like workflows, custom agents, and the retired product label. The current consumer workflow is named Work. OpenAI also documents workspace agents for organizations that build recurring, governed workflows. Check which product a guide describes before following its menus or limits.
Choose Chat, Work, Codex, or cloud browser
Start with the kind of outcome you need.
- Chat
- ask a question, search, brainstorm, summarize, or get a quick conversational response.
- Work
- research a topic, analyze information, or create a finished document, spreadsheet, presentation, report, or Site.
- Codex
- inspect a repository, write or debug code, run tests and commands, or review software changes.
- Cloud browser
- let Work navigate and enter information on supported public websites when a connected app cannot complete the task.
How to use ChatGPT agent workflows now
If an older guide tells you how to use ChatGPT agent, translate its Agent mode step to Work. Availability depends on the paid plan, rollout, and workspace permissions.
- Open ChatGPT on web, mobile, or desktop and select Work where it is available.
- Start a new Work chat or open a Project that already contains the relevant files and instructions.
- Describe the outcome, intended reader, required inputs, constraints, and final format.
- State which sources, apps, files, or websites Work may use. Exclude anything it does not need.
- Define actions that require your approval and conditions that should stop the task.
- Follow progress, answer questions, and change direction if the task drifts.
- Review the finished deliverable, sources, calculations, and proposed actions before relying on it.
Use a bounded task brief
A strong brief replaces a vague request such as research our market with a checkable assignment. Name the goal, source boundary, output, approval rules, and completion test. Keep permissions proportional to the task. The clearer the boundary, the easier it is to detect missing work or an incorrect assumption.
- Goal
- the concrete result to produce.
- Inputs
- files, links, apps, and facts it may use.
- Constraints
- market, time period, budget, tone, exclusions, and required sources.
- Output
- format, length, structure, and destination.
- Approval
- actions that require confirmation before they happen.
- Stop
- ambiguity, missing data, sensitive access, or cost that should pause the task.
- Review
- facts, totals, links, recipients, and acceptance criteria to verify.
Example task for ChatGPT Work
Create a two-page competitor brief for our onboarding product. Use only the five linked product pages and their current pricing pages. Compare audience, core promise, onboarding steps, and price. Put unknown facts in a questions section. Contacting people, signing in, and publishing are outside the task. Return a source-linked document and a five-point review checklist.
Good tasks for Work
- Research a bounded question and create a source-linked report.
- Analyze supplied files and build a spreadsheet with stated assumptions.
- Turn approved material into a presentation or structured document.
- Monitor a defined source for a change and report only when a condition is met.
- Prepare a comparison, itinerary, or plan that a person reviews before action.
When ordinary Chat or a person is a better fit
Use Chat for a short answer, open-ended thinking, or a quick rewrite. Keep a person in control of high-stakes medical, legal, financial, employment, account, and public-communication decisions. A broad instruction such as handle my inbox or manage the business gives too much scope. Split it into a narrow task with limited data and an explicit approval gate.
How cloud browser changes the old agent-mode workflow
Cloud browser lets Work use supported public websites when an app cannot complete the task. OpenAI says the current launch version cannot accept credentials, sign in, use autofill or password managers, or complete payments. It may also stop at a CAPTCHA or blocked site. Include the exact site and desired outcome, then review its links and screenshots before using the result.
Review every consequential action
Before approving an action, confirm the site, recipient, item, date, price, account change, and exact commitment. Keep passwords and payment details out of the chat. Enable only the context needed for the current task. Stop if the system reaches the wrong site or uses the wrong information. A confirmation screen reduces risk; it does not prove that every earlier assumption is correct.
Review the result like delegated work
- Compare the final output with the task brief and required format.
- Open important sources and verify dates, numbers, names, and quoted claims.
- Check formulas, totals, filters, and missing rows in generated spreadsheets.
- Inspect every recipient, permission, price, date, and destination before approval.
- Record unresolved questions instead of accepting a confident guess.
- Revise the brief when the same failure repeats in later tasks.
Frequently asked questions
A ChatGPT agent was the product name for a supervised system that could plan and perform multi-step tasks with tools. In the current ChatGPT product, Work is the replacement for longer tasks and finished deliverables.
Agent mode ChatGPT instructions in older guides describe the retired interface. Select Work for current multi-step workflows. Use cloud browser for supported public-site actions and Codex for software development.
Open Work, provide a bounded task brief, add only the necessary context, monitor progress, answer questions, and review the deliverable. Old instructions that say to select Agent mode from the plus menu no longer match OpenAI's current help page.