OpenAI dots: how an agent that works without being asked actually works

Where this comes from

Dots are available on paid ChatGPT plans (Pro and Business Premium) and in beta for enterprises, outside the European Union, Switzerland and the UK.

What a dot is, in one sentence

A dot is an agent with a name, a face, memory and its own cloud computer, which you direct by messaging it in ChatGPT, by voice, in Slack or in Teams. It takes on a goal that lasts weeks, like keeping a sales proposal current or turning customer complaints into fixes, and decides on its own when to get back to work.

The brain isn’t new: it is GPT-6 Astra, the same model that runs Codex. OpenAI itself says the technical novelty is limited. What changes is everything around the model.

What changes is who starts the next turn

Everything around the model has a name: the harness. It is the loop, the tools, the memory, the environment and the brakes. There are three shapes today, and they differ on a single question: who starts the next turn?

1. Chat assistant. You start every turn. Without you, nothing happens.
2. Session agent. You open the task and it runs a think, act, check loop inside it until it delivers. That is Codex and Claude Code.
3. Always-on agent. You, a clock, an event or the agent itself starts the turn. That is dots, Meta’s Muse and open projects like OpenClaw.

The piece OpenAI calls new is the time budget: a setting that guides how long a dot spends on a goal, alongside a clock that can wait and helper agents that run in parallel.

The five ways a dot wakes up

• You message or call it, in ChatGPT, Slack, Teams or by voice.
• It schedules itself. The dot decides when to pause and when to come back, with no fixed timetable.
• A recurring task, with a time zone and an end date, like a daily calendar review.
• An app event, like a new comment on a document, delivered by signed webhook (the standard called MCP Events).
• Proactive research while you’re away, with read-only tools only: it can read and take notes, but it can’t send messages or change anything.

The tricky one is the second trigger. When the model books its own return, “why did it wake up now?” gets harder to answer. Open-source projects prefer a fixed system heartbeat, precisely because the user is the one paying for the tokens.

Three layers of memory and a computer that stays on

A dot remembers in three ways: the current conversation, ChatGPT’s general memory and its own notes on preferences, decisions and ongoing work. Notes carry across every channel.

Here is the clearest gap: today you can’t view, edit or delete an individual note. To make a dot forget something, you reset the whole agent. For companies under GDPR or Brazil’s LGPD, that matters.

The cloud computer keeps files, software and browser sessions between uses. It doesn’t inherit your VPN or your logins. On supported sites, you type the password into a secure form that goes straight to the machine’s browser, and the model never sees it. For code, the dot opens a task in Codex; for research and files, in ChatGPT Work.

The brakes: your rules, a separate reviewer and a monitor

Every action that touches an account or data goes through two barriers before it happens.

1. Your rules. For each kind of action you pick: act without asking, act if you already asked for it, ask first, or hand it to you.
2. A second model that checks. Auto-review looks at an email’s recipient and content, for example, against your instructions and the safety rules. It lets it through, asks for approval, hands it back to you or blocks it.
3. A monitor over everything, which can pause or stop the dot, and an activity view where you follow along and change course.

Some guards never come off: changing passwords and moving money always stay with you. And OpenAI itself warns that rules are instructions the system can still misapply, and that stopping a task doesn’t undo what was already done.

The takeaway is clear: outside attacks barely get through. The bigger risk is the agent overstepping on its own, and it grows the longer the chain of tasks.

What it costs and who pays

A dot has no price of its own. It comes with ChatGPT’s Pro and Business Premium plans, and chatting with it doesn’t count against chat limits. But the heavy work it hands off to Codex and ChatGPT Work burns those products’ allowance, and the “deeper work” quota isn’t public.

In practice, the dot is the showcase that pulls customers up to the pricier plans, launched the same day. The announced next step is selling more dots and more speed for a flat fee.

What this teaches anyone who wants agents in their operations

Dots are a good product and will teach the market what to ask for: an agent with an owner, clear rules, initiative and a brake before acting. At the same time, they run only on OpenAI’s infrastructure, with a single model provider, no way to install them in your own environment and memory you can’t delete item by item.

For many companies, that is enough. For anyone handling sensitive data, legacy systems or regulation, the same design has to live in-house.

• An AI Analyst maps which routines need an always-on agent and which work fine on demand.
• A Forward Deployed Engineer builds the agent inside the customer’s environment, with logged triggers, a reviewer before external actions, deletable memory and whichever model makes the most sense.
• Training teaches the team to write rules, read the activity log and approve safely.

Support, finance, procurement and sales are just examples. If it’s a task that repeats, it can be automated, and now that includes “keeping an eye on things”.