If you are looking for ChatGPT's agent mode, you are too late. OpenAI's help page now says ChatGPT agent is no longer available and points to ChatGPT Work, launched on July 9, 2026. Anthropic went further two months later: on September 16, Claude Cowork and regular chat started merging into a single Claude. For a small business wondering whether these tools can replace a custom AI agent, the question has changed.
The agent is no longer a hidden feature you switch on to try. It sits in the tool your staff open every morning: a Work tab next to Chat at OpenAI, the conversation itself at Anthropic. The window that used to answer questions can now read a folder, fill in a form, and keep working while its user sits in a meeting.
Here is what the two agents do today, according to the vendors themselves, where they break, and when a task you hand them should become a real process.
The short answer, for the busy
As of September 2026, ChatGPT's agent mode no longer exists: OpenAI replaced it with ChatGPT Work, an agent launched on July 9, 2026 that runs multi-step tasks and hands back documents, spreadsheets and presentations. Claude Cowork, available only on paid plans, has been merging into regular Claude chat on the Pro and Max plans since September 16, 2026; Team and Enterprise plans keep the two modes separate for now. Both agents read files, browse the web, use your connected apps, can take control of the computer, and run scheduled tasks. They are good at one-off research and file wrangling. They break on website logins, long unattended runs, and anything that has to be repeatable and auditable. And every file they open is processed on the vendor's servers: when those servers are outside Quebec, section 17 of Quebec's private-sector privacy law (Law 25) applies.
What happened to agent mode, and to Cowork
The names have moved faster than the articles about them.
At OpenAI, ChatGPT is now split into two experiences: Chat, for quick questions, and Work, which OpenAI describes as an agent for longer, multi-step work and finished deliverables. The old agent mode, which you started by typing /agent and which was capped at 40 requests a month on the Plus plan, is gone. Work draws on your plan's included usage instead, and a complex task uses more of it. On web and mobile, Work is available on eligible paid plans (Plus, Pro, Business, Enterprise, Edu). In the desktop app, OpenAI opens it to every plan, including Free, with limited access.
At Anthropic, Cowork was a separate space for long tasks, built on the same architecture as Claude Code. On Pro and Max accounts that have received the new version, there is no "Cowork" button any more: you describe what you want in any conversation and Claude decides whether to answer or get to work. The rollout is gradual. Team and Enterprise plans, the ones a small business normally uses, keep the two options separate for now, and Anthropic promises Enterprise admins at least 30 days' notice before anything changes.
Anthropic says its users found it frustrating to decide where a task belonged. OpenAI keeps two tabs but puts the agent one click away, even in the free desktop app. For a small business, that convenience has a direct consequence: every employee's chat tool can now act.
What these agents actually do
Both products offer the same four capabilities.
1. Read and produce files. In the desktop app, the agent works in the folders you open to it, reorganizes them, writes a report from them, and hands back a spreadsheet with working formulas or a presentation you can edit in PowerPoint.
2. Use your connected apps. Gmail, Google Drive, Microsoft 365, Slack: once connected, the agent pulls what it needs mid-task. Anthropic's documentation describes an order of preference: connector first, browser second, screen as the last resort. Reading Slack through its connector takes seconds; navigating Slack on screen is slow and brittle.
3. Browse and act on the web. ChatGPT Work has a cloud browser, a remote computer with its own sessions that keeps going after you close your laptop. Claude uses a browser built into its desktop app, or your own Chrome.
4. Take over the computer and work alone. Both can click, type and navigate your desktop software (in beta at Anthropic, on Pro and Max). Both also run scheduled tasks: summarize the inbox every Monday, or update a document whenever a supplier emails.
Where they are good
What the good uses share: one person defines the task and one person judges the result. Preparing for a meeting with a prospect. Putting three supplier quotes side by side in a table. Sorting hundreds of scanned invoices. Turning meeting notes into a deck.
Mistakes are cheap because someone reads the output before using it. The agent replaces an afternoon of shuffling files, not judgment, and it is a sensible first step toward AI that takes action.
Where they break
The vendors write these limits down themselves.
Logins. OpenAI warns that some websites block automated agents even when the same site works fine in your own browser. The agent also stops at every sign-in, two-factor code or confirmation. Your payroll provider's portal may well be that kind of site.
Unattended runs. Anthropic advises against scheduling tasks that touch sensitive files, send messages on your behalf or make purchases. The reason is prompt injection: a booby-trapped email or web page can carry instructions the agent mistakes for yours. Both vendors have added automated checks, and both acknowledge those checks do not eliminate the risk. Anthropic puts it bluntly: you remain responsible for every action Claude takes on your behalf.
Repeatability. This is the least visible limit. A task handed to a desktop agent lives in one person's account, with that person's connectors, folders and wording of the day. When that person goes on vacation, the task goes too. Anthropic even notes that local Cowork session history is stored on the employee's computer, and that admins cannot centrally manage or delete it.
Law 25: the agent sees what you show it
When Claude takes over the screen, it takes screenshots to find its way, so it can see anything displayed there. Anthropic states that work done in a cloud session, including local files opened through the desktop app, is processed on its servers rather than staying on your computer. OpenAI says the same thing differently: messages and task context may be stored in the cloud even when the work runs locally.
Section 17 of Quebec's Act respecting the protection of personal information in the private sector requires a privacy impact assessment and a written agreement before personal information is communicated outside Quebec. The employee file the agent opens, the customer spreadsheet it sorts: each one goes to the vendor's servers, and becomes a communication outside Quebec as soon as those servers are elsewhere. Neither vendor's help pages promise processing in Quebec. We walk through the mechanics in our article on AI agents running on US servers.
Then there is the account type. By default, OpenAI does not use content from Business and Enterprise workspaces to train its models; on personal accounts, it is a setting the user has to turn off. And ChatGPT Work comes, in limited form, with the free desktop app. An employee can point an agent at a company folder from a personal account without anyone knowing. That is shadow AI with hands. Hence the case for a managed team subscription, and we compared the Copilot, ChatGPT and Claude offers. On Claude's Team plan, Cowork and cloud sessions are on by default, and it is the account owner who can switch them off.
The third time: when a task should become a process
Here is the working rule we suggest. The first time an employee hands a task to a desktop agent, it is exploration. The second time, it is a habit. The third time, it is a business process that doesn't know it yet, and four questions apply. Are the inputs always the same kind? Does the output always need the same shape? Do other people depend on it? Will you need to show, six months from now, what was done and why?
If so, the task has outgrown the desktop agent. The vendors offer a middle step: on Business and Enterprise, OpenAI provides workspace agents that can be shared with the team, scheduled, and triggered through an API. For a simple flow that stays inside the vendor's tools, that can be enough.
A custom AI agent answers a different need. It runs a defined process for the whole company rather than for one person. It plugs into your systems, including the ones with no off-the-shelf connector. Its rules are written down, versioned and logged, and the path personal information takes is designed from day one, with the assessment Law 25 requires. That is the distinction we draw between an AI agent and ChatGPT: one waits to be spoken to, the other is part of the workflow.
Where to start
Let your teams try ChatGPT Work or Claude, but on a company account, on tasks where someone reviews the result, and without opening customer or employee files to the agent until the section 17 assessment is done. Then, a month from now, ask them which tasks they have handed to the agent three times. That list is the best starting point for a process agent.
That sorting is what we do in a first conversation: what stays with the desktop agent, what deserves a process, and what it costs.
→ Let's talk about your repetitive tasks
Written byXavier Peich