What are ChatGPT dots? OpenAI's always-on agent explained

On September 29, 2026, OpenAI announced that it is adding “dots” to ChatGPT. They are always-on agents powered by GPT-6 Astra. The official explainer page, “Meet dots”, describes dots as “agents that keep work moving and only reach out to you when a decision is needed”. Each dot has its own computer and browser in the cloud, and keeps working even when your own computer is switched off. It handles everything from research and preparing documents to data analysis and software development, and uses your past conversations and preferences in its work.

How this differs from before#

The biggest difference is that work moves forward between conversations. Until now, ChatGPT returned an answer when it received a question or request, and the work ended there. The ChatGPT agent that appeared in 2025 also operates a virtual computer, but it acted only within the single task it was asked to do.

What you hand to dots is an ongoing responsibility. A dot tracks progress, works out for itself what to do next, and keeps working as things change. It also decides when to pause and when to resume, so you don’t have to set up a schedule for every check. Multiple pieces of work can be split among separate agents running behind the scenes and done in parallel, and you can keep talking to the dot in the meantime.

The official pages list examples of how to hand work off. For preparing an external event, you hand over the plan and your emails with the venue, and it updates the headcount as RSVPs come in, and checks with you before asking for a revised quote. For a sales proposal, it revises the numbers whenever the customer’s requirements or seat count change, and follows up on the security reviews still outstanding. For feedback on software, it investigates reported bugs, prepares fixes and routes them to code review before merging. Each is work that you keep adjusting as circumstances change.

VentureBeat framed the change as “from copilot to someone you hand whole jobs to”. The same day, OpenAI also announced “ChatGPT Space”, where people and agents work in the same documents.

Where you talk to a dot, and memory#

You can talk to a dot from the ChatGPT desktop and mobile apps, Slack, Microsoft Teams and by phone. Wherever you talk to it, it is the same dot, and memory isn’t reset when you change places. Messages stay in each place, and the dot uses the context it needs across places. Before passing something it learned in a private conversation to someone else, it checks whether that is allowed. The first one is created in the desktop app or the browser, and mobile browsers aren’t supported.

In Slack, you can message the dot directly, or call it by mentioning it in a channel you’ve added it to. But simply adding it to a channel doesn’t start monitoring. You need to tell the dot which events in which channels to watch.

Memory has two layers. One is the conversation context: messages, instructions, materials and the results of tools it ran. The other is notes the dot writes itself, recording preferences, decisions and work in progress, carried into the next conversation. These notes are separate from ChatGPT’s memory, and the official explanation says they aren’t a record of everything you’ve said.

Where the work happens: the cloud and your own computer#

A dot’s work basically runs on its own computer in the cloud. You can open that computer’s screen at any time. When you want to operate it yourself, you choose “Take over” to take control, and “Return control” to give it back.

You can also connect one of your own computers, which lets the dot use local files, code and apps. In that case, the computer needs to be online with the ChatGPT app open. Plugins such as Gmail, Google Drive and GitHub work if they are connected and the requested action is permitted. The announcement says plugins connect to more than 4,000 apps.

What stands out is that permissions are separated by what you connect. Connecting Slack or Teams doesn’t reach email, other apps or your computer. Connecting your computer doesn’t let it read all of your past ChatGPT or Codex conversations.

When a website needs a login, the dot sends you a login request. You enter your credentials in an input field that opens separately from the conversation. The cloud browser has a separate session from the browser on your own computer. For sites that refuse access from the cloud, you can ask it to try on the connected computer.

How to hand off work, and how to stop it#

The official guide recommends starting with work that needs regular checking or updating, looking at the first results, fixing what’s missing, and then widening the scope. For work that repeats at set times, you tell the dot what to check, when to run it (including the time zone and an end date), what kind of changes to tell you about and where to send the results, and have it save that as a schedule. If the connected service supports it, you can also ask for event-triggered work, such as investigating when a bug report arrives in a particular Slack channel.

You check on delegated work in “Activity” in the dot’s profile in the desktop app, and check repeating schedules in “Scheduled”. There are three ways to stop it.

• “Pause” stops only the main work the dot is on right now. Other work you handed off and scheduled runs don’t stop
• Individual pieces of work you’ve handed off are opened and stopped in “Activity”
• Repeating schedules are disabled or deleted in “Scheduled”

With any of these, actions already carried out aren’t undone. Even deleting the dot itself doesn’t undo changes made in connected apps or messages already sent.

How permissions and safety work#

With an agent that acts on its own, the question is what you can hand off, and how far. According to the official explanation, a dot works inside three mechanisms from the start: built-in safeguards, the permissions you’ve already given each app in ChatGPT, and an automated review.

Before actions that affect your account or share information elsewhere, the automated review checks them against your instructions, permissions, custom rules and safety rules. The result decides whether to go ahead, wait for your approval, or leave the action to you. For example, a password change is done by you. How the scope of an instruction is treated is also specific: asking for a draft reply doesn’t mean you have allowed sending it.

Even while you haven’t asked for any work, a dot checks connected apps and looks for ways to help. The tools used for this are read-only, and can’t send messages, change the content of apps, or operate the browser or computer.

When you want finer control, use custom rules. You write what the action is and choose one of four treatments.

• “Take action without asking”: run it without confirming
• “Take action when you say so”: run it only when you clearly ask, and otherwise confirm just before running
• “Ask before taking action”: ask for approval before running
• “Hand off to you”: leave the action to you

The official examples set sending a message to a customer to “Ask before taking action”, and deleting a shared project file to “Hand off to you”. But the official page itself writes that custom rules are instructions the dot tries to follow, and it can get them wrong. Rules can’t grant access to apps or computers, or remove built-in safety rules.

On data handling, ChatGPT’s data settings also apply to conversations with dots and to the work they do. OpenAI says it doesn’t directly train models on a dot’s self-initiated research or on a dot’s notes.

Plans and pricing#

The eligibility on the official page is as follows. All are staged rollouts, and even on an eligible plan it may not appear right away.

• Pro 100, Pro 200, Pro 500: users aged 18 or over, outside the European Economic Area (EEA), the UK and Switzerland
• Business Premium: rolling out worldwide
• Enterprise: rolling out worldwide. It is off by default, and a workspace admin turns it on

The first dot is included in Pro and Business Premium plans at no extra charge. Conversations with dots don’t count toward ChatGPT usage limits, but work that dots start or manage in Work or Codex counts toward the limits of those products. The plans include an allowance for deeper work, and for one month from launch that allowance is expanded. The announcement says it will later let you add dots and raise a single dot’s speed and monthly workload.

Pro 500 is what the OpenAI Help Center describes as a new Pro plan at $500 a month. The “new $500-a-month plan” some press reports mentioned appears to refer to this plan. Dots also work on Pro 100 and Pro 200.

Enterprise admins turn on “Use dots (Beta)” under “Permissions & roles” in workspace settings. According to the Help Center, adding to Slack and Teams, access to local computers and use of custom rules can each be managed through separate settings.

What to think about when using it in your company#

You design the scope of what you hand off. There are several layers of safety mechanisms. Even so, the official pages themselves write that custom rules can be wrong, and that stopping a dot doesn’t undo completed actions.

A reasonable approach is to hand off only reading and drafting at first, and keep actions with outside effects, such as sending and deleting, at “Ask before taking action” or “Hand off to you”. Then widen the scope as you see results. In Enterprise, one option is to allow local computer access and custom rules for only some users at first. The more apps you connect, the wider the dot’s reach, so limit connections to the apps you need.

For an example of an agent using an unexpected means to get a job done, the access OpenAI’s agents made to a United Nations site is also worth a look.