ChatGPT Agent: What Replaced Agent Mode

ChatGPT agent is no longer available, verified against OpenAI's help pages on October 26, 2026. Start in ChatGPT Work for longer tasks and finished deliverables; cloud browser handles supported website steps inside Work. If you relied on recurring agent tasks, inspect your current schedules before assuming the work still runs.

ChatGPT Agent Is Retired: What Actually Changed

OpenAI's ChatGPT agent help page states:

ChatGPT agent is no longer available.

The notice points to ChatGPT Work, the experience for delegated tasks and finished outputs, and cloud browser in ChatGPT, its tool for supported website workflows. It does not state a retirement date. October 26, 2026 is the verification date for this article, not a claimed shutdown date.

For a founder trying to recover a weekly supplier report, the practical change is the starting point: select Work and describe the report. For an operator whose old agent task ran on a schedule, the starting point is the current task list. Finding a replacement capability does not establish that an existing job transferred successfully.

What Is Agent Mode in ChatGPT?

Agent mode was ChatGPT's way to perform multi-step online work using tools, rather than only describe how you could do it. Its virtual browser could navigate websites and fill forms; connected apps supplied research data; users could take over the browser and schedule repeat work. Those are historical capabilities documented on the agent help page.

Think of Work as the person you delegate a finished assignment to, and cloud browser as one of the instruments available to complete it. You choose the assignment. ChatGPT chooses whether an app connection, browser interaction, or both fit the supported steps.

Where the Old Capabilities Live Now

The old bundle maps to several current capabilities, with Work as the place to start. This is a capability map, not a guarantee that every old task, login or schedule migrated.

The Operator section of the retired agent page documents that earlier integration. It is part of the product's history, not a reason to look for another Operator switch.

For an operations lead, this distinction prevents a common mistake: forcing browser clicks for a task that an authorized app can handle directly. OpenAI's cloud browser instructions explicitly say ChatGPT may use an app or plugin instead.

Which Plans Get Work and Cloud Browser?

On web and mobile, Work requires eligible paid access; cloud browser explicitly excludes Free and Go. Keep the eligibility language when making a rollout decision:

• Work: The Work product page lists Plus, Pro, Business, Enterprise and Edu for web and mobile, and says all plans on desktop. The Help Center describes web/mobile access as eligible paid plans and qualifies desktop access by plan and workspace. It also names eligible ChatGPT for Healthcare users for event-triggered Work tasks. Check the account and workspace, rather than assuming every surface has identical access.
• Cloud browser: OpenAI says paid ChatGPT plans in supported regions, excluding Free and Go. Rollout and workspace permissions can affect availability. It runs within Work, so an eligible subscription alone does not override a workspace restriction. Cloud browser availability.

Work follows Codex's usage structure, with included usage and additional credits; consumption varies by task. The retired agent page's monthly request allowances are not the current Work budget. Check the current Work usage guidance before sizing recurring work.

How to Use ChatGPT Agent Workflows Now

ChatGPT Work is the current starting point for a delegated deliverable. Begin with a bounded job whose result you can inspect, such as a supplier comparison with source links and unresolved questions.

The current Work setup and cloud browser flow support this sequence:

1. 1

   Select Work and define the output

   Open ChatGPT on web or mobile, select Work, and start a conversation. Name the deliverable, relevant sources and review criteria. For example: “Compare the suppliers on this shortlist. Create a spreadsheet with availability, source links and unanswered questions. Ask before contacting anyone.”

2. 2

   Provide only the context the assignment needs

   Add the relevant files or authorized app context. For a website task, name the exact site and account context. Let ChatGPT choose a supported app or cloud browser route.

3. 3

   Handle sign-in through the secure flow

   Approve website access when asked. Use the secure sign-in form and complete any authentication steps there. Do not paste passwords or security codes into the conversation.

4. 4

   Inspect the outcome and unresolved steps

   Check the deliverable against its sources. Review confirmation requests and identify anything the site blocked or that needs a manual final step. Ask for corrections before making the workflow recurring.

Where Is Agent Mode in ChatGPT?

The retired mode is not a missing control you need to enable. Look for Work instead. If Work is unavailable in a managed workspace, ask an admin to check Workspace settings > Permissions & roles, including Work Cloud or Work Local and the separate browser controls. OpenAI documents these distinctions in its Work access settings.

A browser login on your laptop does not sign in the cloud browser. It has its own sessions. If you need to intervene, ask ChatGPT to take over the browser; the current takeover instructions say it will provide a link.

What to Do With Old Scheduled Agent Tasks

Inspect Scheduled before rebuilding anything. The current help pages describe task management, but do not establish that every old agent schedule automatically migrated or continues to execute. An existing task should be judged by its current settings and results.

Open Scheduled in ChatGPT. You can also go to Settings > Notifications > Manage tasks. If the desktop app does not show Scheduled, use the web version. These are the current routes in OpenAI's scheduled-task documentation.

For a recurring competitor update, use this recovery sequence:

1. Inspect the existing task. Read its instructions, schedule and recent results. Check whether it is paused or waiting for approval. Confirm that any required app account remains authorized.
2. Preserve the useful specification. Record the intended output, sources, timing and review owner before changing the setup. If a functioning task is already present, avoid creating a duplicate.
3. Run the work once in Work if it needs rebuilding. Confirm that today's sources and access produce the intended result. A successful old run is not evidence that a changed workflow still works.
4. Review, then schedule. Create or revise the recurring task only after the output is acceptable. Check the next result before treating the replacement as an operational dependency.

There is a specific file-access constraint worth checking before you rebuild: a scheduled task created in a Project cannot access uploaded files or files stored in that Project, according to the current task instructions. Putting the task beside a spreadsheet does not make that spreadsheet a usable scheduled source. Confirm the task's supported inputs before relying on it.

Current Task Limits Are Not Agent Message Limits

The current limit is an active-task count, separate from Work usage and model access:

These counts come from OpenAI's current scheduled-task limits, checked October 26, 2026. Other eligible plans depend on the account or workspace. Free and Go appearing here does not grant access to Work's cloud browser.

For an operator, task capacity and execution capacity are different questions: having room to create another task does not promise sufficient usage or compatible sources to finish it. Pause unused tasks before rebuilding a duplicate, and check the specific job's access and output.

When Work Fits

Work fits when the assignment ends in something you can review: a researched memo, a comparison spreadsheet or a presentation. It is a sensible next step for a founder who previously delegated that kind of work to agent mode. Keep quick questions in Chat; OpenAI positions Codex for software development and technical tasks. Choosing an experience.

For the deeper adoption decision, read the ChatGPT Work review. For a wider shortlist organized around the job, use the best AI agents guide. The migration decision here is narrower: recover the outcome you needed, with current access and a clear reviewer.

What Is Overhyped About the Replacement?

A new home for the capability is not proof of uninterrupted automation. The migration map tells you where to start; it cannot establish what happened inside your account. Treat old schedules, permissions and outputs as things to inspect.

Remote execution does not mean every job finishes unattended. Cloud browser can keep working after you leave, but it pauses for sign-in, input or confirmation. Websites can block automated access, and some steps remain unsupported. Build a manual fallback for a deadline-sensitive booking or record update. Cloud browser limitations.

Website permission is not transaction approval. OpenAI separates access to a site from approval for consequential actions. Letting Work read a supplier portal should not be your operating instruction to place an order. Name the permitted result and the point where a person must decide. Website access and confirmations.

For a CTO, the useful adoption test is whether the result is inspectable, the sources are accessible and the failure path has an owner. A polished deliverable alone does not establish that a recurring process is dependable.

Frequently Asked Questions

Is ChatGPT Agent Free?

ChatGPT agent is retired. Its cloud-browser successor within Work is available on eligible paid plans, excluding Free and Go. Basic scheduled-task access is a different entitlement and does not restore agent mode. Current browser availability.

What is the difference between a bot and an agent?

Bot is a broad term for automated software. An AI agent can use a model and tools to choose and carry out steps toward a goal. Work's role is delegated task execution; a conversational answer alone does not establish that an action happened.

Can you provide some examples of agentic AI?

Examples include asking Work to assemble a sourced report, delegating a supported signed-in website task to cloud browser, or creating a task that monitors for changes. The tools and permissions available to the account determine what can run. Work capabilities.

Is OpenAI agent Builder free?

OpenAI introduced Agent Builder as included with standard API model pricing, so that did not mean free model execution. It is a separate developer workflow tool, and its current documentation marks it deprecated. It does not restore retired ChatGPT agent mode.

What is the role of an agent in ChatGPT?

Work handles longer assignments that require multiple steps and a finished output. The retired agent mode performed online tasks with a virtual browser and other tools. Use the current Work experience for a new assignment. Chat and Work roles.

How much does ChatGPT agents cost?

There is no current agent-mode allowance to buy. Work access depends on an eligible paid plan, and usage follows the structure described in OpenAI's current Work guidance. Do not budget from the retired mode's message limits. Work access and usage.

What are the big 4 AI agents?

There is no standard set of four that decides this migration. Compare agents by the work you need done, the systems they can access and how you will review the output, rather than treating a marketing shortlist as a technical category.

What is the most advanced AI agent right now?

There is no universal winner established by this capability check. For an operator, the useful test is whether an agent can complete the specific assignment with acceptable access, review effort and failure handling.

Is ChatGPT an agent or LLM?

ChatGPT is an application that uses large language models, or LLMs. An LLM supplies language and reasoning capabilities; an agent combines a model with tools and a process for taking steps. Work is the agent experience for delegated tasks within ChatGPT.

What are the risks of using AI agents?

OpenAI names prompt injection, phishing and unintended actions among cloud-browser risks. Prompt injection means malicious content tries to redirect the assistant's instructions. Limit access to the task, inspect important results and use the secure sign-in flow. Safeguards do not eliminate every risk. Cloud browser safety guidance.

Your Monday Move

If you owned recurring agent work, inspect its task list and latest output first. If you are starting fresh, give Work a bounded deliverable with named sources and a human reviewer. A workspace admin should resolve access before a team builds a dependency; a builder should keep software work in its appropriate development workflow.

Wait on unattended use when the required site or input does not work. If you only used Chat for questions, this retirement does not require adopting Work. The immediate action is to validate one useful assignment, then decide whether it deserves a schedule.

For practical updates on AI workflows and product changes, join the newsletter.