OpenAI unveiled Dots during DevDay, a personal artificial intelligence agent designed to continue working on tasks after the conversation with the user ends. Rather than merely answering a question or carrying out a single command, Dots follows projects, monitors the arrival of new information, updates its work, and then presents the results for review.
Powered by GPT-6 Astra, Dots operates inside a dedicated cloud computer for each agent. It can use the browser, create files, and run tools. Users can also connect their personal computer to enable it to work on spreadsheets stored on the device or development tools installed on it. OpenAI says it isolates cloud environments and restricts the scope of code access in an attempt to limit the effects of incorrect commands or malicious code.
Ongoing tasks instead of separate commands
OpenAI offers examples such as monitoring the notes of an application’s users, identifying recurring issues, preparing and testing fixes, and then presenting videos explaining the changes to the developer. Dots can also select clips from interview transcripts to produce short videos, and prepare episode notes and drafts for social posts. However, these are use cases provided by the company and do not prove that the agent will achieve the same level of performance on every task.
Communication with the agent is not limited to the ChatGPT window; users can send messages or have a voice conversation, while Dots continues working through Slack and Microsoft Teams. The extensions ecosystem enables connections to more than 4,000 applications, with the scope of available information and tools remaining tied to the permissions granted by the user.
Memory and approvals determine the level of autonomy
Dots uses ChatGPT memory and recent conversations, and conversations with it can add information to that memory. However, users currently cannot view, edit, or directly delete the individual pieces of information the agent remembers. Disconnecting an application prevents it from obtaining new information, but does not erase what it has already learned; the help page says that deleting Dots’s private context requires deleting the agent itself, while files and memories transferred to ChatGPT are managed separately.
The ability to work continuously does not mean that Dots can take any action automatically. Proactive research is limited to reading permitted sources and creating private notes, and during that process it cannot send messages, modify application content, or control the computer. Actions such as sending an email or editing a file require authorization and additional security checks.
The Auto-review system evaluates instructions, custom rules, and security requirements before carrying out planned actions. Users can set explicit restrictions, such as prohibiting email sending. Purchases require approval, while actions such as changing passwords and transferring money between accounts remain in the user’s hands. OpenAI acknowledges that errors can occur, so work automation does not eliminate the need to review results, particularly when they involve sharing information or affecting accounts.
Availability and the enterprise version
Dots is being rolled out gradually in eligible markets to ChatGPT Pro and Business Premium users, while Enterprise users can try it in a beta version when enabled by their workspace administrators. The first agent comes at no additional charge with the Pro and Business Premium plans, with usage limits for extensive tasks. Dots conversations do not count toward ChatGPT limits, but Codex and ChatGPT Workspace tasks that it starts or manages count toward their own limits. Initially, only one agent can be created, while the ability for multiple agents to collaborate is part of future plans.
OpenAI is also offering specialized agents for companies in a limited enterprise preview, with identities, access data, and tools suited to their responsibilities. The company is working on Microsoft Agent 365 integration to enable these agents to be managed and monitored through the governance and security tools already in use at organizations.
What changes in practice?
Dots represents a shift from a model in which the assistant waits for every request to one in which an ongoing responsibility is delegated to an agent that follows it over time. However, the value of this model will not be measured only by the number of connected tools, but by its ability to distinguish between what it can carry out and what requires a human decision, and by how clearly users can control memory and permissions. The product’s current limitations, such as the inability to manage individual memories directly and availability restricted to specific plans and markets, remain important constraints when evaluating its use for sensitive tasks.