ChatGPT agent mode for business: what it does now, and where it falls over.

What is ChatGPT agent mode, and does it still exist?

ChatGPT agent mode was the setting that let ChatGPT work through a multi-step task on its own, using a browser and your connected apps. OpenAI has since retired it. Its help centre now says ChatGPT agent is no longer available and points people to ChatGPT Work for longer, multi-step tasks and finished deliverables.

So if you are searching for the agent mode button and cannot find it, that is why. The idea behind it is still there, in a bigger form. ChatGPT Work is built for tasks with a clear outcome, like a brief, a deck, an analysis or a file you can review, rather than quick answers. That is how OpenAI's own Work guide describes it.

Not everyone thinks the swap is like for like. On OpenAI's community forum, users have said Work does not reproduce the way agent mode handled their workflows, and an OpenAI support reply in September acknowledged the transition had disrupted people. If you had a routine built on agent mode, expect to rebuild it rather than flick a switch.

What can ChatGPT Work actually do for a business?

Work takes a job you describe, plans the steps, pulls context from the apps you connect, and hands back a finished file. It can run a task once, on a schedule, or as a monitor that watches for changes. On the desktop app it can also use your local files, apps and a browser.

The pieces that matter for an owner:

• Plan mode. Before it starts, ChatGPT asks questions and drafts a step-by-step plan. You change it or approve it. This is the single most useful habit for anyone new to agents.
• Scheduled tasks. A task can repeat, monitor or refresh something over time. Cloud tasks run without your computer being awake; tasks that need local files need the computer on and the app running, per OpenAI's scheduled tasks docs.
• Plugins. Connections to email, calendars, Slack and other tools, so it can read the context it needs. Which plugins you get depends on your plan and workspace settings.
• Credits. Longer or more complex tasks use more of your plan's allowance. Check your usage before you hand it something big.

It sits on the paid ChatGPT plans on web and mobile, and in the desktop app across plans. If you are still choosing a plan, our ChatGPT training covers how to set a team account up so staff use it on real work.

Which business jobs suit an agent, and which are still demos?

Agents do well on jobs with a clear finish line, a small number of sources and a human checking the result: research briefs, weekly reports, tidying a spreadsheet, a first draft of a proposal. They struggle with long chains of logins, jobs where a wrong click costs money, and anything that needs your own rates and rules.

We sit in on a lot of owners' first attempts. This is roughly how the jobs sort out.

The last row is the one most people miss. If the steps are the same every time, you do not need something that thinks; you need something that runs. Our n8n vs Zapier vs Make comparison covers those tools, and our AI automation work is mostly that kind of job: fixed steps, with AI used only where a judgement call is needed, like reading an invoice.

Why do AI agents go in circles?

Agents loop when the goal is vague, when they hit something they cannot get past, like a login, a pop-up or a changed web page, or when nothing tells them what "done" looks like. They keep trying variations instead of stopping. A clear finish line, fewer steps and a rule to stop and ask fixes most of it.

It is the most common complaint we hear from workshop graduates who start building agents at home: it makes mistakes, or goes round in circles. Nearly every time, the task was too big for one run. What helps:

• Say what done looks like. "A one-page table of the five cheapest suppliers, with links" beats "find me a better supplier".
• Split long jobs. Gathering, drafting and checking work better as separate steps. A West Perth accounting firm we trained, AustAsia Group, set up three agents for exactly this: one gathers a client's information, one starts the draft, one reviews.
• Tell it when to stop. OpenAI's own guide suggests limits like telling Work to stop before sending anything. Use them.
• Give it your material. An agent with your price list, past quotes and a written process makes far fewer wrong turns than one guessing from the open web.

Is ChatGPT Work safe to use with business data and accounts?

It can be, with the right setup. Work asks you to approve a plan, lets you approve significant steps and redirect it mid-task, and on business plans admins control who can use it, which plugins it connects to and whether it can browse. The risk is giving an agent more access than the job needs.

Our rules for any agent, ChatGPT or otherwise:

• Narrowest access that works. OpenAI's docs say the same for scheduled tasks: start with the least access that lets the task succeed. Read-only first; add send or edit rights once you trust it.
• A person approves anything that leaves the building. Emails to customers, payments, orders and anything posted publicly.
• Company accounts, not personal ones. Business and Enterprise workspaces give the owner control. Staff running agents on personal logins with company data is the risk we see most.
• Be careful with "full access" on the desktop. OpenAI warns that background tasks in full access mode can change files and run commands without asking.

If your data is regulated, like health, NDIS or legal files, decide where it is allowed to go before you connect anything. Our security page sets out how we handle data on every build.

ChatGPT Work, Manus or a custom agent: which should a business use?

Use ChatGPT Work when your team already lives in ChatGPT and the jobs are research, reports and drafts. Look at other general agents, like Manus, if you want to compare. Build a custom agent or automation when the job runs every day, touches your core systems, or needs to work the same way every time.

General agents are good at variety and poor at repetition. A custom build is the reverse. Our Manus AI review covers where that tool beats ChatGPT and where it does not. For the jobs that genuinely need to run unattended across Xero, your job app and your inbox, a custom AI agent built around your process is usually the more reliable choice, because it only does the steps you designed.