OpenAI launches Dots, persistent agents with their own cloud computer

OpenAI presented Dots at its DevDay event on September 29, 2026. The company described the new agents as “visual bubbles” that can act autonomously within the ChatGPT platform. According to OpenAI’s press release, Dots are the first generation of persistent agents that have their own cloud execution environment, allowing them to stay active between user sessions.

Each Dot is powered by the GPT‑6 Astra language model, the latest version of OpenAI’s architecture. In addition to the model, the agent receives a dedicated virtual computer and a web browser, both hosted on OpenAI’s cloud infrastructure. This combination enables the agent to run code, browse web pages, and store data locally within its own environment, without relying on the invoking user’s session.

Dots appear in the ChatGPT interface as animated circular avatars that can express simple emotions through color and motion changes. OpenAI states that the visual design is intended for users to perceive the agent as a collaborative entity rather than a simple text tool. The graphic representation also serves as a hook for the agent to request permission before performing actions that involve external data access.

How Dots’ architecture works

In practice, each Dot receives a lightweight virtual machine that includes CPU, memory and storage allocated dynamically based on workload. OpenAI has noted that the environment is isolated from other agents, ensuring that information processed by one Dot does not mix with that of another. The integrated browser allows the agent to open pages, fill out forms, and extract structured information without direct human intervention.

One distinctive feature is persistence: the agent’s internal state, including memories of prior interactions and temporary files, is retained as long as the virtual computer remains active. Thanks to this persistence, Dots can perform background tasks such as price monitoring, calendar updates, or periodic report generation, and notify the user when the result is ready. OpenAI indicates that continuous execution is subject to time and resource consumption limits defined by the subscription.

Applications and plugin ecosystem

OpenAI has made more than 4,000 applications available through its plugin marketplace, and each Dot can install and use any of them autonomously. Plugins include productivity services, databases, e‑commerce platforms, and third‑party APIs. When an agent needs, for example, to check a bank account balance or publish an article in a CMS, it opens the corresponding plugin and carries out the operation within its secure environment.

Dots are also designed to learn user preferences over time. By analyzing past interactions, the agent can anticipate the preferred format, tone, or level of detail, and adjust its responses without the user having to specify each time. OpenAI clarifies that this learning occurs within the agent’s isolated environment and that users can delete or reset the history at any moment.

Practical limitations

Although OpenAI markets Dots as autonomous agents, the company emphasizes that human oversight remains essential. ChatGPT Pro users can set execution time limits, review activity logs, and revoke plugin permissions in real time. Moreover, the service price is tied to cloud resource consumption; OpenAI has warned that costs can rise quickly if agents run intensive processes or maintain prolonged sessions.

The initial rollout of Dots is limited to ChatGPT Pro subscribers, who will be able to activate one or more agents from their account settings. OpenAI has promised to expand availability to enterprise plans in the coming months, but has not confirmed an exact date for general access. In the meantime, users who enable Dots will be billed based on CPU time and bandwidth used by each agent.

What OpenAI clarified on October 2

A privacy and safety FAQ published by OpenAI on October 2 says Dots share memory with ChatGPT and may retain context drawn from conversations and plug-ins for as long as the agent exists. Disconnecting a service stops new access but does not erase information already absorbed into that context, and users cannot yet inspect, correct, or delete an individual memory. Proactive research remains read-only: it cannot send messages, change content, or control a browser. The most sensitive actions, including changing a password or transferring money, require the person to take over.

With Dots, OpenAI aims to transform how users interact with AI, shifting from single queries to continuous assistants that can manage complex tasks without constant intervention. In practice, ChatGPT Pro subscribers now have a tool that combines cloud processing, web browsing, and access to thousands of applications, all under the control of a visual avatar. The true impact will depend on how each user balances the agent’s autonomy with the need for supervision and the cost associated with prolonged use.