GuidesAI at work
Creating an agent in ChatGPT: ChatGPT Work and workspace agents
Creating an agent in ChatGPT: agent mode is retired and replaced by Work, ChatGPT Business workspace agents, plugins, prices and limits.
Contents
- What has replaced ChatGPT agent mode?
- How do you hand a task to ChatGPT Work, step by step?
- How do you create an agent for the whole team in ChatGPT Business?
- How much does an agent in ChatGPT cost?
- What breaks when the agent goes into production?
- What Balt changes for a team that builds its agents in ChatGPT
- Where to start?
To create an agent in ChatGPT in November 2026, the first thing to know is that agent mode no longer exists: OpenAI has retired it and points to ChatGPT Work, which carries out long, multi-step tasks and can run at a set time or on a trigger. For an agent the whole team shares, ChatGPT Business and Enterprise offer workspace agents, and for reusable instructions, plugins replace custom GPTs, which are retired on 11 December 2026.
This guide follows OpenAI’s help centre as it reads on 2 November 2026, on a small-business task. It describes what replaces agent mode, how to hand a task to Work, how to create an agent for the team, what connected apps allow, what it all costs and what breaks. What happens to custom GPTs is covered separately, in our guide to custom GPTs in your company.
What has replaced ChatGPT agent mode?
Agent mode, launched in July 2025, drove its own browser to complete a task on the web, with monthly limits: 40 messages on Plus, 400 on Pro and none on the free plan. As of 2 November 2026, OpenAI’s help page states that it is no longer available, and its functions, along with those of custom GPTs, are now split across four building blocks that serve different purposes and come with different plans.
| Building block | What it does | Who it is for |
|---|---|---|
| ChatGPT Work | carries a long task through to the deliverable, once, at a set time or on a trigger | eligible paid plans, one person |
| The cloud browser | browses on a remote computer during a Work task, even with your machine closed | paid plans, except Free and Go |
| Workspace agents | a shared agent, scheduled, triggered through an API or used in Slack | ChatGPT Business, Enterprise and Edu |
| Plugins | instructions, reference files and apps you can reuse | every plan, in place of GPTs |
The first question is therefore one of ownership: a task one person carries out for themselves belongs in Work, an agent the whole team uses belongs in a workspace agent, and know-how to be reused belongs in a plugin.
How do you hand a task to ChatGPT Work, step by step?
OpenAI presents ChatGPT Work as an agent designed for long, multi-step work that ends in a finished deliverable: a document, a spreadsheet, a presentation, a report or a website. Take a seven-person training company that receives its registrations by email and wants an up-to-date table of sessions and remaining places every Monday. The example is built for illustration, but it brings together what Work does best: reading, consolidating and producing a file.
Step 1: describe the deliverable
You describe the result rather than the steps: “A table of this week’s registrations, by session, with the name, the company, the stated funding and the remaining places, plus a list of registrations where a piece of information is missing.” Work breaks the task down, carries it out, shows its progress and asks its questions along the way, which you answer in the conversation.
Step 2: connect the mailbox
Apps are connected from the settings: Google Drive, SharePoint, Microsoft Teams, Outlook for email and calendar, Slack or GitHub. For the training company, Gmail or Outlook is enough, and the permission level is chosen app by app, from “Always ask” to “Allow all actions”, with read-only and low-risk actions in between. For a first trial, read-only covers everything the task needs.
Step 3: schedule or trigger it
A Work task can run once, repeat on a schedule, react to a trigger or watch for a change. On the Plus, Pro, Business, Enterprise and Edu plans, triggers include a new Gmail email, a message in a Slack channel where ChatGPT has been added, or a pull request on GitHub. The Monday table thus becomes a weekly task, and each new registration can also trigger an update.
Step 4: approve the important actions
Work asks for approval before important actions, and the cloud browser asks for confirmation before a booking or a payment. You sign in to sites through a secure form without the model seeing the password, and access to each site can be set to “Always ask”, “Auto approve” or “Always allow”. Some sites block automated browsers, which usually comes to light on the day the task fails.
How do you create an agent for the whole team in ChatGPT Business?
In ChatGPT Business, Enterprise and Edu, the workspace agent is the agent the team shares. You create it from the Agents tab, choose its model, add tools such as Google Calendar, Google Drive, Slack or SharePoint, custom MCP servers, web search or image generation, and then test it before publishing. Every version is kept and can be republished, which lets you roll back after a change to the instructions.
Once published, the agent can be shared, used in Slack, scheduled or triggered through an API. Its write actions are set to “Always ask” by default, and its creator can choose “Never ask” or a custom rule. On Enterprise, workspace agents are turned off by default, and the administrator decides who can create them.
For each app, the agent connects either with the account of the person using it or with an account that belongs to the agent. In the second case it has the same permissions whoever is talking to it, which is the question an agent’s identity raises: should two people with different permissions get the same answer?
How much does an agent in ChatGPT cost?
OpenAI states that Work follows the same usage structure as Codex: an allowance included in the plan, then credits. Workspace agents are paid for in credits, and the ChatGPT Business Standard seat costs $20 per person per month on annual billing and $25 month to month, with two seats minimum and no official price in euros. Plus costs $20 a month for one person.
The cost that appears on no invoice is the time spent writing the instructions, testing the agent on real cases and rereading what it produces in the first weeks. For a task like the registrations table, it adds up to hours at the start and then minutes a week, provided you decided from the outset who rereads what.
What breaks when the agent goes into production?
The retirement of agent mode, a little over a year after its launch, shows first of all that an agent configured with a vendor lives at the pace of that vendor’s decisions: what was built for one model or product often breaks with the next. Keep the instructions and the description of each task in a company document, so that the next migration starts with a copy.
The second risk is prompt injection. An agent that reads emails and web pages reads text written by strangers, and OpenAI itself once gave the example of a booby-trapped comment pushing the agent to retrieve a reset code from Gmail; today it writes that its protections “do not eliminate every risk”. The only serious defence is to make sure the agent cannot do anything serious even when the injection gets through, as our article on injection hidden in a CV explains.
The third is approval. The “Allow all actions” and “Never ask” levels exist, and everyone ends up using them on the day confirmation starts slowing the work down, whereas the rule we apply cannot be adjusted: what leaves the company goes through a person, and that rule is only worth something if it is built into the tool rather than written into the instructions.
What Balt changes for a team that builds its agents in ChatGPT
Balt is an AI employee of the company rather than an agent the team builds: you write to him in Teams, Slack, WhatsApp, on the phone or in his web app, he works in more than 3,200 connected tools, and what he learns with one person becomes a skill for the whole team. The Monday table becomes a team routine, one that reports back even when nothing happened.
Nothing he prepares for a client or a supplier goes out without your approval, with no setting to lift it. He costs from €99 a month for 40,000 credits shared by the whole team, with unlimited users and 20,000 credits offered at sign-up, and his database is hosted in France and the European Union. For open-ended writing, deep research or code, ChatGPT goes further than he does, and the full comparison is on our ChatGPT Business or Balt page.
Where to start?
With a single task whose deliverable can be reread in a minute, handed to Work in read-only mode for three weeks before you open up any write action. If it serves the whole team, rebuild it afterwards as a workspace agent, deciding before you start which account it connects with and who approves what goes out. The testing method is in our protocol for testing an agent.
Frequently asked questions
Does ChatGPT agent mode still exist?
No. As of 2 November 2026, OpenAI’s help centre states that ChatGPT agent is no longer available and points to ChatGPT Work for long, multi-step tasks and finished deliverables. The browser the agent used has been replaced by a cloud browser, available in Work on paid plans, except Free and Go.
How do you create an agent in ChatGPT?
On your own, you hand a task to ChatGPT Work, which carries it out in several steps, once, at a set time or on a trigger. As a team, in ChatGPT Business or Enterprise, you create a workspace agent from the Agents tab, add apps and instructions, test it and then publish it. For reusable instructions, plugins replace custom GPTs, which are retired on 11 December 2026.
How much does an agent in ChatGPT cost?
Work follows the same usage structure as Codex: an allowance included in the plan, then credits. Workspace agents in Business and Enterprise are paid for in credits, and the ChatGPT Business Standard seat costs $20 a month on annual billing and $25 month to month, as of 2 November 2026.
Can a ChatGPT agent send emails on its own?
It can if you let it. Write actions for workspace agents are set to “Always ask” by default, but their creator can choose “Never ask”, and each connected app can be set as far as “Allow all actions”. Work also asks for approval before important actions, bookings and payments included.
Sources
- OpenAI, ChatGPT agent (help centre, accessed 2 November 2026)help.openai.com
- OpenAI, ChatGPT Work (help centre, accessed 2 November 2026)help.openai.com
- OpenAI, Cloud browser in ChatGPT Work (help centre, accessed 2 November 2026)help.openai.com
- OpenAI, ChatGPT Workspace Agents for Enterprise and Businesshelp.openai.com
- OpenAI, Introducing workspace agents in ChatGPT (22 April 2026)openai.com
- OpenAI, Apps in ChatGPT (help centre, accessed 2 November 2026)help.openai.com
- OpenAI, Custom GPT retirement and migration FAQhelp.openai.com
- OpenAI, ChatGPT Business pricing (accessed 2 November 2026, in French)openai.com

