OpenAI’s dots are AI agents that keep working when you log off 🤖

Cliquez ici pour lire en français

OpenAI is introducing dots, autonomous agents designed to help users with their everyday tasks and projects. According to the company, they are powered by GPT-6 Astra, have their own computer in the cloud, can run 24/7, and gradually learn from their users’ preferences and feedback.

Agents built to work on their own 🤖

Dots are meant to change how people use AI. Instead of simply answering one-off requests, they can take on tasks, follow projects and keep working in the background.

Each dot has its own identity to manage its access and permissions. It also gets a cloud computer and browser, which lets it write and test code when needed.

You can also open your dot’s computer at any time to review its work. And if you choose, you can let it connect to your own devices, such as a laptop, so it can help directly in your working environment.

An AI that learns how you work 🧠

Dots are designed to evolve with use. The more you work with your agent, the more it is supposed to learn about your preferences, how you think, your goals and your standards. In theory, you can keep handing it new tasks or ideas without starting each project from scratch.

A dot can follow several pieces of work in parallel while keeping their context. OpenAI says this continuity should help turn ideas and projects that never got the time or resources into tasks that can move forward step by step.

What dots could do in practice 🚀

OpenAI offers several use cases to show how dots work. In app development, for example, a dot can spot recurring requests in customer feedback. It can then scope small improvements, fix bugs, build and test the changes, and submit pull requests (proposed code changes) along with videos showing what changed.

Similarly, during a product launch, a dot can keep track of the target audience, the positioning and the creative requirements. If the product’s scope shifts, it can analyze the consequences, revise the launch materials and prepare proposed edits. You can still review the work and approve the changes.

For researchers and content creators 🔬

OpenAI’s examples also cover scientists and content creators. In research, a dot can follow a scientific question, analyze new data, rerun analyses and look into unexpected results. It can then update an article’s figures and revise the explanations when the conclusions change.

Content creation works in a similar way. A dot can learn your tone and what your audience expects. From a new interview transcript, it can pick out passages that could become clips, prepare show notes and suggest social media posts. It can also carry your edits across the different formats and keep learning from your feedback.

Available in ChatGPT, SMS, Slack and Teams 📱

Dots are built to follow you across several channels. You can use them in ChatGPT on the web, on mobile and on desktop. They can also message you to report progress on a task, ask a question or put a decision to you.

They also work with SMS, Slack and Teams. You can start a project in ChatGPT, add a detail by text, then give more context in Slack.

Context carries over between channels, so the dot can keep working without losing what you’ve already told it. Pro users can also join a limited waitlist to talk to their dot through iMessage and RCS (the successor to SMS).

You stay in control 🔐

Despite their autonomy, dots come with control mechanisms. OpenAI says they rely on ChatGPT’s existing protections, with extra safety and privacy safeguards tailored to the tasks they can perform.

You can set limits, manage permissions and follow progress. You can also step in when a decision needs your approval.

Each dot works on its own cloud computer. Your personal computer and its contents therefore stay separate, unless you choose to connect your device.

Guardrails on what agents can do 🛡️

On compatible sites, dots can use saved passwords without exposing them to the model. When working in the background, they can also run « proactive research. »

In that mode, they draw on apps you have already connected, using read-only tools. They cannot send messages, edit the content of your apps or control your browser or computer.

OpenAI also says it has built in safeguards to detect potentially dangerous behavior and to defend dots against certain malicious instructions, a risk often called prompt injection. If a safety issue is detected, the monitoring system can pause or stop their work.

Rules for when to act and when to ask ⚙️

You choose which apps your dots can access and manage permissions from existing ChatGPT settings. Dots also come with built-in rules that decide when they can act on their own and when they must ask for approval.

Custom rules let you allow certain actions, require approval for them or block them. In every case, the built-in safety requirements still apply. You can follow the work done, including background tasks, from the Activity view, and redirect your dot when needed.

Checks before sensitive actions 🔎

OpenAI says it also uses an automatic review system to check actions that could affect accounts or share information.

This system checks whether actions match your instructions, your custom rules and the safety requirements. It is meant to decide which tasks can continue, which need approval and which must be done by you directly.

Some sensitive operations, such as changing a password, always stay with you. OpenAI also acknowledges that dots can still make mistakes, so work with significant consequences should be checked.

Specialized dots for businesses 🏢

OpenAI is also planning a version for organizations. Unlike the personal dot, a specialized dot can be given specific responsibilities inside a company. Each organization can assign it its own identity, credentials and the access it needs for its tasks. Over time, it is said to adapt to new information and improve through team feedback.

OpenAI’s first tests reportedly focus on procurement, invoice processing, email marketing, customer service and commercial contract management. For these pilots, OpenAI’s teams work directly with organizations to define each dot’s responsibilities, the tools it can access, and how its work is checked and approved.

A planned integration with Microsoft Agent 365 💼

OpenAI is also working with Microsoft to plug specialized dots into the enterprise governance and security tools of Microsoft Agent 365. The goal is to let companies manage these agents with the Microsoft tools they already use.

OpenAI says dots are rolling out in ChatGPT for Pro and Business Premium users in eligible markets. Enterprise users, including Edu and Healthcare, can access a beta once their workspace admin turns it on.

The first dot is included in Pro and Business Premium plans at no extra cost. It can be used 24/7 to chat, think things through and follow the tasks you give it. Plans also include an allowance for deeper tasks, with extended limits during the first month after launch.

Eventually, OpenAI plans to let users add several dots and increase their capacity, either by speeding up their pace or by raising the volume of tasks they can handle each month.

A new way to organize work with AI 🤖

Conversations with a dot do not count against ChatGPT’s usual usage limits. However, when you ask your dot to launch or manage tasks in Codex or ChatGPT Work, those tasks count against the usual limits of those services.

To get started, you create your first dot from the ChatGPT desktop app or a computer browser, then connect the apps you want it to access. After that initial setup, you can keep messaging it from the ChatGPT mobile app.

With dots, OpenAI is betting on agents that can work over time, keep context, learn from feedback and handle several stages of the same project. The company says it will keep improving them as it learns which uses make them genuinely useful day to day.

Do you think autonomous agents like these can really change how professionals use AI every day? Tell us in the comments 👇