# OpenAI Puts Always-On Agents Inside ChatGPT With Their Own Cloud Machines

Dots run on GPT-6 Astra, continue working after the user closes the app, connect to more than 4,000 applications, and launch first for Pro and Business Premium subscribers.

- Published: 2026-09-30T06:34:52.927Z
- Canonical: https://polylog.news/ai/2026-09-30/openai-puts-always-on-agents-inside-chatgpt-with-their-own-c
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI News](https://openai.com/index/introducing-dots), [Polylog editors](https://polylog.news)

OpenAI's second DevDay announcement changes the basic unit of work in ChatGPT from the conversation to the agent. [Dots](https://openai.com/index/introducing-dots) are persistent assistants that retain context, hold ongoing responsibility for a project, and keep running between sessions. Each dot gets its own cloud computer, so it continues working after the user closes their laptop.

The specifications matter more than the framing. Dots run on GPT-6 Astra, have their own browser, and can [connect to more than 4,000 applications](https://t.me/aipost/8308). Users assign each dot a name and a role, then attach mail, calendar and messaging accounts. A dot can be messaged from ChatGPT, Slack or Teams. Alongside dots, OpenAI introduced [ChatGPT Space, a shared workspace where people and agents work on the same material](https://t.me/ai_machinelearning_big_data/11043). The rollout starts with ChatGPT Pro and Business Premium in eligible markets, with an administrator-enabled beta for Enterprise, Edu and Healthcare workspaces.

The open question is security. A long-running agent with its own browser, credentials to thousands of third-party services, and no human reviewing most of its steps is the same configuration that produced OpenAI's own agent incidents this year at Hugging Face and on Australian government websites. OpenAI says dots are designed to keep the user in control, but it has not published the containment architecture or the permission model in the detail an enterprise security team would need before granting a dot access to a company mailbox.

## What this means

The mechanism is distribution combined with persistence. An agent that survives the session and holds credentials becomes the system of record for a workflow, which raises the cost of switching away from it far above the cost of switching chat interfaces. Exposed are the agent-framework startups that sell orchestration, scheduling and memory as separate products, and the enterprise administrators who must now write access policy for non-human identities inside Slack and Teams. What matters here is distribution: OpenAI is shipping this persistent runtime to a user base it puts at 1.2 billion weekly users. The likelier outcome over the next few months is that enterprise adoption is limited by identity and audit controls rather than by model quality, and that OpenAI publishes a permissions and logging specification because procurement teams will require one. This would be proved wrong if administrators enable dots broadly without new controls, which would indicate the security concern matters less than the productivity gain.

## What to watch

- Whether OpenAI publishes a per-dot permission and audit-log specification. That is the prerequisite most enterprise security teams require before granting mailbox and calendar access.
- The first reported incident involving a dot acting on a third-party connector without authorization. That would test how much containment the cloud computer actually provides.
- How agent-framework vendors respond, since a persistent runtime built into ChatGPT removes a large part of what they currently sell.