OpenAI Dots: the AI that keeps working after you close ChatGPT

On Tuesday, September 29, OpenAI introduced something that seems more important to me because of the shift in dynamics than because of the product name.

It's called Dots, and the idea is simple to explain: you stop talking to an AI only when you need something and start giving it responsibilities that stay alive even after you close the conversation.

A Dot can receive an objective, connect to the tools you authorize, and keep making progress while you're doing something else. OpenAI describes it as an "always-on" agent, available 24/7 and built on GPT-6 Astra. The official announcement confirms that each Dot has its own cloud computer and can connect to more than 4,000 applications through the plugins ecosystem.

That's where the interesting shift is.

From asking for tasks to delegating responsibilities

Until now, much of our use of ChatGPT works in turns.

You open the app.

You ask for something.

You get the result.

You make corrections.

You close it.

With Dots, OpenAI wants certain tasks to have continuity.

You can tell it what objective you're pursuing, which applications it can consult, what things it can do on its own, and which ones must go through your approval first.

The difference seems small written that way, but in a workflow it changes quite a bit.

A developer could leave it the responsibility of reviewing user comments, detecting repeated reports, and preparing fixes for review.

Someone in sales could ask it to keep a proposal up to date while the client's requirements change.

A content creator could connect it with transcripts and have it prepare clips, notes, and drafts as new material appears.

These are examples OpenAI uses directly in its Dots announcement.

The important word here is continuity.

It has its own computer

Each Dot initially works inside a computer hosted in OpenAI's cloud.

It doesn't need control of your personal computer to function.

It can use its browser, work with connected applications, and keep several projects active. You can also open that computer and check what it's doing.

Access to your own machine exists, but it's optional and starts turned off. According to OpenAI's documentation, if you decide to enable it, the Dot can work with local files, start Codex or ChatGPT Work tasks, and use your browser when necessary.

That already places Dots quite far from the traditional chatbot.

Here's how the flow changes

To me, this table explains Dots better than simply calling it “a new agent.”

More than 4,000 apps

Another strong point is in the connections.

OpenAI claims that Dots can use its plugin ecosystem to connect with more than 4,000 apps.

They can also follow the same conversation from ChatGPT, Slack, or Microsoft Teams without losing the project context. OpenAI is even preparing support through text messages.

This matters because an agent without access to external tools ends up being an assistant that knows a lot but can do very little.

When it can check documents, calendar, email, development tools, or enterprise systems, its usefulness changes.

The risk changes too.

The real problem will be deciding how much access to give it

An agent available 24/7 with access to work information needs fairly clear limits.

OpenAI built in several layers of control precisely for that.

During what it calls proactive research, the Dot can review information from the apps you've already connected, but it uses tools restricted to reading. In that state, it can't send messages, modify content, or control the browser or computer.

For other actions, there are rules.

You can allow them automatically, block them, or indicate that they require approval.

Some sensitive actions remain mandatorily under human control. Changing a password is one of the examples OpenAI gives.

There's also an automatic review system before certain actions that could modify accounts or share information.

That doesn't mean we can connect twenty services and forget about it.

The company itself acknowledges in its documentation that Dots can make mistakes and recommends reviewing any work with important consequences.

And I think that warning deserves quite a bit more attention than the spectacular part of the launch.

An agent that learns how you work

Dots also tries to solve another common problem with AI: having to explain several times how you want things done.

OpenAI says the agent learns through feedback and retains context about your goals, preferences, and standards.

In theory, after working with you long enough, it should know what result you consider acceptable.

That can be especially useful in repetitive processes.

If you constantly correct the same type of detail in a document, proposal, technical task, or post, the expectation is that the Dot will incorporate that feedback.

Not because it has developed human judgment.

Because it has persistent context about how you've worked with it.

That difference matters.

Who can use it?

The rollout began for ChatGPT Pro and Business Premium users in supported markets.

In Pro, the initial launch excludes the European Economic Area, Switzerland, and the United Kingdom. Enterprise, including Edu and Healthcare, has access through a beta that must be enabled by the administrator.

OpenAI also notes that the rollout is gradual, so having one of those plans doesn't necessarily mean the button will appear immediately.

The first Dot is included in Pro and Business Premium at no additional cost.

Conversations with the Dot also don't consume the usual ChatGPT limits, although tasks you launch within Codex or ChatGPT Work still count against their respective quotas.

These details are covered in the official setup guide.

This feels a lot like having a digital teammate

This is where I personally see the most interesting part.

For years we used AI assistants as a tool that waits for instructions.

Agents are turning that relationship into something more like delegating work.

You still define the goal and make the delicate decisions, but you no longer necessarily have to describe every intermediate step.

For someone who develops software, for example, I find it much more appealing to say:

"Take care of reviewing these reports and bring me the cases that really need my attention."

Than spending half an hour every morning manually going through all the sources.

The same applies to client follow-up, administrative tasks, content, research, or documentation.

The savings aren't just in the AI doing a task quickly.

They're in not having to remember to start that task every time.

It also raises the cost of making mistakes

The more autonomy a tool has, the more important its permissions become.

A mistake inside a chat can end up as an incorrect answer.

A mistake by an agent connected to external systems can turn into an incorrect action.

That's why I think sensible Dots adoption should start with reversible tasks and limited access.

• Reading before writing.
• Preparing before publishing.
• Proposing before executing.
• Asking for authorization for any sensitive action.
• Periodically reviewing which applications are still connected.

Later you can expand responsibilities as you get evidence that the workflow works.

I wouldn't start by handing over the keys to everything.

My takeaways

Dots shows quite clearly where OpenAI wants to take ChatGPT.

Fewer isolated sessions.

More persistent work.

The agent has its own computer, browser, context, access to thousands of applications, scheduled tasks, and the ability to keep working while you do something else.

But the truly important part is going to be control.

If these agents become reliable enough within well-defined limits, they can remove an enormous amount of repetitive work.

If we set permissions without thinking, we can also multiply our mistakes.

The technology is already starting to have autonomy.

Now it's time to learn to delegate it well.

Conclusion

OpenAI Dots doesn't catch my attention for having a new avatar inside ChatGPT.

It interests me because it turns something we've used mainly on demand until now into a persistent process.

You give it work.

You leave.

And the work can keep moving forward.

That model has much more potential than simply getting a slightly better answer to our next prompt.

It also requires that we treat permissions, reviews, and limits as part of the work, not as a secondary setting.

References

• OpenAI: Introducing dots
• OpenAI Help Center: Getting started with your dot
• OpenAI: ChatGPT Release Notes
• Reuters: OpenAI takes on Meta with always-on Dots agent