In the morning, you open your computer and the work is done. OpenAI introduced dots agents

Chatbots have so far worked simply: you write something, they reply, and wait for the next instruction. OpenAI has now introduced OpenAI dots at its DevDay developer conference – agents that don’t wait. They receive a goal and access to applications from you and continue working on it, even if you have ChatGPT long closed. In practice, it should look like this: in the morning, you open your computer and instead of a list of unfinished tasks, their in-progress or even completed version awaits you.

The company is responding to Muse, Meta’s personal AI agent, whose rocket launch in America we wrote about last week. But dots are aimed elsewhere – instead of restaurant reservations, they primarily handle work, i.e., code, presentations, analyses, or business proposals. We’ve reviewed the official announcement, security documentation, and help, so below you’ll find what agents can do, how much they cost, and most importantly, how much you can trust them.

The arrival of a dot agent is similar to a new employee starting

Regular ChatGPT functions as a smart consultant: you ask, you get an answer, and its work ends there. A dot is more like a new colleague who takes over a task and sees it through to completion. OpenAI itself compares its onboarding to a new employee starting. It receives almost everything you would give a person on their first day of work – its own identity for managing access and permissions, a computer, and a browser.

Introducing dots, powered by GPT-6 Astra.

Remarkably capable, always-on agents built to handle everything. pic.twitter.com/inY5BZj3Kz

— OpenAI (@OpenAI) September 29, 2026

Each dot has its own virtual computer in the cloud with Linux and a Chrome browser, on which it browses the web, analyzes data, creates files, and, if necessary, writes and tests code. You can look into its computer at any time, monitor what it’s doing, and intervene in the work yourself if needed. You can also connect your laptop to it, but this option is disabled by default.

The agents’ brain is GPT-6 Astra, OpenAI’s most capable model, which the company introduced in early September. It is supposed to excel mainly in sticking to the scope of the assignment, and if it’s unsure, it asks for clarification. For an agent working unsupervised for hours or days, this is a crucial feature. Dots connect to applications via a plugin ecosystem, and the list already includes over 4,000 services, according to the company.

The third difference is memory. A dot takes over the memories ChatGPT has stored about you and creates its own – even from connected applications. The longer you work together, the better it should know your preferences, way of thinking, and what constitutes a job well done for you. TechCrunch reminds us that most of these capabilities already existed in Codex and similar tools. The main novelty is that OpenAI has packaged them into a single agent that acts on its own initiative.

As if you were talking to a colleague

Communication should resemble writing with a colleague. You text or call with a dot in ChatGPT on the web, on your computer, and on your mobile. For subscribers in the USA, text messages via iMessage and RCS will even be available, but they must first sign up for a waiting list. The agent itself sends ongoing progress reports, questions, or points on which you need to decide.

In addition to ChatGPT, a dot also works in Slack and Microsoft Teams. You can start a project in ChatGPT, add details in Slack, and the context will transfer between all channels. OpenAI itself warns that these go through an external provider, so it’s better not to write sensitive information in SMS messages. You can also connect a personal email, but the dot will not get its own email address yet.

Creation has one limitation: you can only create a dot in the ChatGPT desktop application (including Windows) or in a browser on a computer. You can then use it in the mobile application, but the mobile web is not supported. During setup, you name the dot and choose its appearance – there are characters and pets to choose from, or it can generate one for you. In its profile, you can then see what it has in progress, planned, and completed.

What can a dot do for you?

OpenAI showed five model situations that clearly demonstrate where an agent saves the most time – with work that is repetitive and changes based on new information.

• Developer: a dot monitors customer feedback, fixes minor bugs itself, and tests small improvements. It then submits a completed pull request with a video demonstrating the changes for review
• Product Manager: when the product scope changes, a dot adjusts the campaign, presentations, and other launch materials
• Scientist: with each new batch of data, a dot repeats the analysis, investigates unexpected results, and updates graphs for an upcoming study
• Salesperson: a dot compares client requirements with documentation, builds a prototype of a key integration, and recommends a colleague who has already handled a similar order
• Content Creator: from an interview transcript, it selects moments for short videos, prepares episode notes, and suggests social media posts

All examples share the same principle: the agent prepares, the human approves. Meanwhile, the developer works on a major feature, the salesperson negotiates with the customer, and the scientist is in the lab.

However, a small detail from one of the first testers is the most telling. His dot noticed that the tester forgot to issue an invoice, prepared it itself, and sent it after approval. This is precisely what OpenAI is building the entire product on: the agent is supposed to deliver work before you even think to ask for it. How well this will work outside of carefully selected examples will only be proven by regular use.

But dots are not just for huge companies and their incredibly busy employees. They can also handle daily routines – every morning they go through your calendar and summarize what awaits you, remind you of a deadline, conduct research, or regularly check something. You manage recurring tasks in the dot’s profile, where you set their interval, time, and notification upon completion. You don’t have to switch between threads, as a dot can handle multiple projects at once, and you can continuously add new tasks to it.

It works even if you don’t give it any tasks

This is where dots differ most from classic chatbots. When you’re not actively working with an agent, it searches in the background for ways it could help you – OpenAI calls this proactive exploration. For example, a dot might go through connected applications and notice that your travel plans have changed, or it might take notes from emails and documents for future work.

To prevent useful initiative from becoming a nightmare, proactive exploration has strict boundaries. It only uses tools designed for reading, so a dot cannot send anything to anyone, change application content, or control the browser or computer during this process. It saves its findings in private notes, and if it decides to act based on them, it must pass the same checks as with a regular task. If you don’t want it to work at all, you can pause it with a single button in its profile.

Even such capable agents are limited by certain rules

An agent with access to email, calendar, and accounts can cause significant damage. OpenAI has therefore built several layers of protection around dots and describes them in considerable detail in a separate security document. The foundation is Custom Rules, where you choose one of four behaviors for specific actions.

• the dot performs the action without asking
• it performs it only if you explicitly instructed it to
• it asks you before performing it
• it hands the action over to you

Above your rules stand fixed limits that you cannot override. Permanent data deletion, software installation from an unknown source, or granting new access requires your confirmation every time. Purchases using cards stored with a merchant require your approval. And some steps a dot will not take at all – password changes or money transfers between accounts will always be handed over to you, even if it can assist with the preparation.

The sharing of personal data is also well thought out. The more sensitive the information, the more specifically you must designate the recipient. A dot will only send health data to a specifically named person, such as a particular doctor. For email or phone numbers, a category (“any airline”) is sufficient, which you can expand with your own rule. Consent applies only to the given task and is not extended by subsequent delegation.

Furthermore, every action that sends data or modifies files is checked before execution by a separate Automatic Review system. For emails, it checks both the recipient and the content to catch a wrong address or information you didn’t intend to share. It runs outside the environment to which the dot has access, so the agent cannot disable or bypass it. Even your consent cannot override basic security requirements.

Passwords are handled by secure login. When a dot needs to access an account, the model pauses, and you fill in the details into a protected form that transmits them directly to the browser without the model seeing them. Stored passwords via an encrypted service work similarly. However, be careful with passwords written in a document or message – the model can read those.

Agents are also protected against so-called prompt injection, i.e., hidden instructions on the web, in an email, or document that try to induce them to do something you didn’t want. Furthermore, the dot’s work is monitored by security monitoring, which pauses it in case of suspicious behavior and notifies you. The progress of all tasks, including those in the background, can be found in the Activity Overview, where you can stop the dot or change its direction.

Can an agent make a mistake?

Yes, and OpenAI states this openly. It always recommends checking work with serious implications, and the help documentation adds that a dot can make mistakes even when adhering to your own rules. That this is not a theoretical threat was recently demonstrated by Meta’s competitor Muse, which, according to The Verge, sent a YouTuber’s address to a stranger.

Also, carefully consider what you entrust to the agent. Disconnecting an application does not erase what the dot has already obtained from it. If you want to clear its memory, you must reset the dot, which will cause you to lose all conversations, memories, and planned tasks. Data from Business, Enterprise, and Edu corporate plans are not used by OpenAI for model training by default; for personal plans, you decide this in the “Improve model for everyone” setting. Notes from proactive exploration are not directly used for training unless they become part of a conversation.

With a corporate plan, they should also be available in the Czech Republic

Dots are starting to roll out now and may arrive at individual accounts in a few days. You get your first dot at no extra charge as part of your plan – provided you have the right one.

The Pro plan costs approximately 4,200 Czech crowns per month (200 dollars), and OpenAI has simultaneously added an even more expensive Pro 500 variant for about 10,500 Czech crowns (500 dollars). However, Pro subscribers in the European Economic Area, Switzerland, and the United Kingdom will not receive dots yet, so Czech users are out of luck. This is all the more unpleasant because, according to Engadget, OpenAI has also simultaneously cut the limits of the two-hundred-dollar plan – for Codex, ChatGPT Work, and messages with the GPT-6 Pro model by roughly half.

The Czech path to dots thus leads through corporate plans. According to OpenAI, the Business Premium package is available in all regions where ChatGPT works (including ours). Enterprise users can try the beta version, but the administrator must first enable it, as it is disabled by default.

Conversations with a dot do not count towards ChatGPT limits, but tasks that the agent initiates in Codex or ChatGPT Work do. During the first month, the dot’s own work also does not count towards the plan’s limits. Later, you will be able to purchase additional dots or increase work speed and volume for a fixed monthly fee – the company has not yet revealed how much this will cost.

Would you entrust a dot with access to your email and calendar?

Sources: OpenAI, The Verge, TechCrunch, Engadget