OpenAI Launches Dots: AI Agents That Keep Working Between Chats

OpenAI introduced Dots at DevDay 2026 on September 29, adding persistent AI agents to its expanding workplace platform. The announcement was part of more than 20 product and feature updates covering ChatGPT, Codex and the company’s models. OpenAI is positioning these agents around ongoing responsibility: users define a goal, and AI continues making progress between conversations.

That changes the question businesses need to ask about AI. Producing a useful answer is one measure of value. Maintaining a project, recognizing when something changes and returning useful work for review is a more demanding test.

What are OpenAI Dots?

Dots are powered by GPT-6 Astra and operate with their own cloud computer and browser. OpenAI says its plugin ecosystem connects them to more than 4,000 applications. Users can communicate through ChatGPT, Slack and Microsoft Teams, with access subject to supported integrations and permissions.

The practical idea is continuity. A person can explain an objective, provide context and establish boundaries, then return later to review progress. The product’s usefulness will depend on how reliably that context translates into the right actions.

What can Dots do?

OpenAI’s examples include turning customer feedback into tested software fixes and pull requests, revising launch materials as product scope changes, and preparing show notes and social drafts from interview transcripts. These are company-described workflows, rather than independently measured productivity results.

For a development team, the interesting possibility is reducing the coordination around smaller changes. An agent that identifies a recurring issue, gathers relevant context and prepares a reviewable fix could remove several handoffs. The engineering team would still need to judge whether the change solves the right problem and meets its standards.

Availability and pricing

OpenAI’s current setup documentation says the first Dot is included with Pro or Business Premium at no additional charge. Deeper work has a plan allowance, with extended limits during the first month. Access is being introduced gradually.

“Included” should not be read as unlimited execution. The launch announcement says conversations with a Dot do not count toward ChatGPT usage limits, while tasks it starts or manages in Codex or ChatGPT Work count toward the applicable limits.

How users get started

Setup begins in the ChatGPT desktop application or a desktop browser. Mobile creation is not currently supported. After setup, users can access their Dot through the mobile app as availability reaches their account.

Local computer access is optional and initially disabled. Users can inspect work in progress, scheduled activity and completed tasks through the Dot’s profile.

This makes a narrow first assignment a sensible evaluation: define the intended output, the information it can use and the actions that need review. The quality of a completed task will be more informative than an open-ended demonstration.

Enterprise controls matter as much as capabilities

For Enterprise workspaces, Dots are disabled by default. Administrators can control access through workspace permissions and roles, with separate controls for local computer access, cloud capabilities and supported messaging channels.

Enabling Dots does not automatically authorize every application. Plugin permissions and each connected service’s authorization continue to constrain access. OpenAI also notes that a Dot’s cloud computer does not automatically inherit a user’s local VPN, browser sessions or device policies.

These details matter when moving from a demonstration to deployment. A company must decide which systems an agent needs, who owns the resulting work and how errors will be reviewed. Giving an agent a broad objective without clear ownership creates an operational problem, even when the underlying model performs well.

Specialist Dots are an enterprise preview

OpenAI also previewed specialist Dots with organizational identities, credentials and defined responsibilities. Initial work is taking place through focused enterprise pilots. The company says it is working with Microsoft on integration with Agent 365 governance and security controls.

The distinction matters: a preview of dedicated organizational agents should not be treated as a broadly available replacement for an entire business function.

How permissions and safety checks work

OpenAI distinguishes proactive research from actions that affect other people or systems. Background research uses restricted, read-only tools. Those tools cannot directly send messages, change connected application content or control a browser or desktop. Follow-up actions must pass the usual rules and checks.

A separate system called Auto-review checks relevant proposed actions against instructions, permissions and safety requirements. Monitoring can also pause work when it detects a concern. Some sensitive steps, including password changes and transfers between financial accounts, must be handed back to the user.

These controls reduce risk; they do not establish that an agent will always make the correct decision. Organizations evaluating Dots should measure unauthorized actions, incorrect changes and missed escalations alongside time saved.

Privacy and memory deserve close attention

OpenAI says Business, Enterprise and Edu content is not used for model training by default. On personal plans, the model-improvement setting governs eligible conversations and work.

There is an important distinction between disconnecting an application and deleting retained context. Disconnecting stops new access, but information already incorporated into the Dot’s context remains. OpenAI’s FAQ says users currently cannot inspect or individually edit or delete specific Dot memories; deleting the Dot clears its own context. Separately stored files and other conversations are not automatically deleted with it.

For businesses handling confidential information, this makes the choice of connected sources consequential from the beginning.

+

The broader direction is toward AI participating in a continuing workflow. OpenAI also announced ChatGPT Space as a shared environment for teams and agents, alongside additional cloud capabilities for Codex. Together, these products indicate an effort to make ChatGPT a place where work is coordinated and completed over time.

From an industry perspective, the opportunity is substantial, but the evaluation needs to be practical. How much useful work reaches review? How often does a person need to intervene? Does the agent recognize when the original objective no longer fits the situation? Does the cost remain reasonable after revisions and failed attempts?

Those questions will determine whether persistent agents earn ongoing responsibility. A polished demonstration can show possibility. Reliable results across weeks of ordinary work will determine adoption.

Frequently asked questions

Is Dots a new AI model?

No. Dots is an agent product powered by GPT-6 Astra. Its role is to combine model capabilities, tools and persistent context for ongoing work.

Can a Dot call me on its own?

OpenAI’s launch setup documentation says Dots cannot initiate calls to users. Texting is described as a limited beta for some U.S. Pro users.

Does enabling a Dot give it access to every company app?

No. Workspace controls, plugin permissions and the connected service’s own authorization govern access.

Can I require approval before actions?

Custom Rules support additional boundaries, but cannot override mandatory safeguards. OpenAI also cautions that Dots can make mistakes when following instructions.

Is proactive research allowed to send emails automatically?

Its research tools cannot directly send messages. Any subsequent action must follow the separate authorization and safety rules.