OpenAI dots: the agent that works while you sleep

Posted on Oct 17 • Originally published at klukyanov.ru

A regular ChatGPT conversation lives as long as the tab is open. Close it, and everything waits for your next message. On September 29 at DevDay, OpenAI showed the opposite: an agent that keeps doing the job after you've gone to bed. OpenAI calls it a dot. Each dot gets its own cloud computer with a browser, access to your apps, and its own memory, and it's on around the clock.

I went through what dots can do at launch, how their permissions work, what they cost and where the limits are. I haven't run a dot myself yet. OpenAI's announcement page returned 403 while I was collecting material, so the facts below come from detailed write-ups that cite OpenAI's announcement and help center. The key numbers match across three or four sources; where there's only one source, I say so.

This is a snapshot as of October 17, 2026. Dots are rolling out gradually, and details will change.

A dot is a persistent AI agent inside ChatGPT, running on GPT-6 Astra (released September 3). Four things set it apart from a normal chat:

Every dot has a name, a look and a handle like @yourname-dot. You can talk to it in ChatGPT on desktop and mobile, in Slack and Microsoft Teams, or call it by voice. Context carries across channels.

The example everyone repeats from the launch: an early tester's dot found an invoice he had forgotten to send, assembled it from an email thread and sent it as a PDF — after he approved it. The last part is the important one.

A dot is a coordinator rather than a worker. It doesn't write code itself; it hands coding to Codex and heavy office work to ChatGPT Work. Those tasks show up as their own threads you can steer directly.

A dot draws on the conversation, relevant ChatGPT memory and its own notes. The notes are the most controversial part of the system: you can't view, correct or delete individual entries. The only full reset is deleting the dot. Disconnecting a plugin stops new access but keeps what was already learned; turning off ChatGPT memory stops sharing but doesn't delete what the dot already received. Credentials, images and screenshots don't go into dot context.

For an agent that reads your mail for months, the lack of a "forget this" button is a real gap.

The core of dots isn't the model, it's the permission system. Under Custom rules, every action gets one of four modes:

On top of your rules sit limits you can't remove:

Pause stops the main task but not necessarily delegated or scheduled work, and nothing is rolled back: a sent email stays sent.

From the GPT-6 Astra system card appendix, as reported in Flavio Copes's deep dive (the only source with the full set):

The third line is the one I care about. In a chain of ten tasks, almost one time in five the agent steps outside its brief. That's not a hack, it's an agent forgetting what it was asked — and long chains of tasks over weeks are exactly how a dot is meant to work. OpenAI says it plainly: "Dots can still make mistakes, so always review consequential work."

The first dot is included, with extended limits for the first month after launch. Post-launch limits, extra dots and extra capacity have no published prices yet. Setup is desktop-only, and users under 18 can't use dots.

This comparison is based on product descriptions, not measurements.

I build apps and agents, so the first thing I looked for was a way to talk to a dot from code. There isn't one: OpenAI hasn't announced an API for dots. It's a ChatGPT product, not a developer platform.

The Codex link is telling, though. A dot can start a task in a Codex cloud environment while your computer is off, send work to your connected machine, or continue an existing Codex thread. OpenAI is building a hierarchy: on top, an agent that knows you and your work; below, specialized executors. If you build agents yourself, selling "an assistant that does everything" just got harder, while a niche executor with a clear interface is still in demand.

A good dot task is recurring, spans several apps and ends in a draft:

A bad task is one where mistakes can't be undone: money, passwords, unchecked emails to important people. Not because the dot is dumb, but because 19.7% out-of-scope on long chains is a lot for irreversible actions. Treat it like a new assistant: drafts first, then widen permissions one at a time.

Shorter weekly write-ups (in Russian) — on Telegram.