OpenAI launches Dots, always-on agents that keep working between conversations

The story

OpenAI has launched Dots, a new class of persistent agents designed to keep working across tools and projects between conversations. Introduced at the company's DevDay 2026 conference in San Francisco on September 29, the product is powered by GPT-6 Astra and represents a shift from AI that responds inside a single chat toward software that maintains context, follows changing work and returns when a result or human decision is needed.

OpenAI's documentation says each Dot operates on its own cloud computer with a browser. It can research, analyze data, prepare documents and build software while a user's device is off. The same agent can be reached through ChatGPT and, when connected, Slack or Microsoft Teams. Switching channels does not create a new agent or erase its memory, although messages remain in their respective channels and the system is supposed to check authorization before carrying private information into a shared setting.

Persistence is the central technical and product change. A Dot can track an assignment as circumstances evolve, decide when to pause or resume, and use background agents to work on multiple tasks. OpenAI gives the example of organizing an offsite: the agent could update attendance as responses arrive and ask for approval before requesting revised pricing. That is a more demanding operating model than completing a bounded prompt because the system must interpret new evidence, preserve goals and recognize when an action exceeds its authority.

The agent draws on the current conversation, relevant ChatGPT memory and its own saved notes about preferences, decisions and ongoing work. It can also use connected plugins within their existing permissions and, if a user separately connects a personal computer, access local files and apps while that machine is online. OpenAI says connecting a messaging channel does not automatically grant access to applications or a computer, an important distinction for organizations mapping where data and authority can flow.

OpenAI says users can create rules governing what a Dot may do and when it must seek permission. Sensitive operations such as changing a password or permanently deleting data require explicit consent, according to Reuters. The company also says business customer data is not used for model training by default, while users of personal plans can opt out of allowing conversations and Dot work to be used for training. Those commitments address familiar concerns, but they do not by themselves demonstrate how consistently the controls will work across thousands of third-party applications.

The rollout is gradual rather than universal. OpenAI's DevDay recap says Dots are available on Pro and Business Premium in eligible markets. Enterprise, Education and Healthcare workspaces can test a beta when an administrator enables it; the feature is off by default in those environments. The documentation also notes that initial creation happens on desktop, and that some mobile support depends on an updated app.

The launch arrived with visible evidence that the experience is not yet frictionless. Reuters reported that several live demonstrations failed to deliver requested voice updates. OpenAI attributed the glitches to the simultaneous rollout of multiple updates. A conference failure does not establish overall product quality, but it is a useful warning against treating polished agent concepts as proof of dependable long-duration execution.

Safety questions are especially important because persistence magnifies both usefulness and error. A conventional chatbot mistake may end with a poor answer. An always-on agent can carry an incorrect assumption forward, act through connected services or expose information in the wrong context. Associated Press reported that OpenAI announced Dots one day after withholding a more advanced model over safety concerns, while chief executive Sam Altman said the company was increasing investment in agent safety, security and monitoring.

INNOVOX analysis: Dots turn continuity itself into a product. The value is not merely that the model can browse or write code; other agents already perform those tasks. The bet is that one persistent identity can remember the user's goals, coordinate tools and return at the right moment without requiring every workflow to be scripted in advance. If it works, the competitive advantage will come from trusted context and dependable follow-through. If it fails, the same persistence could make mistakes harder to notice and permissions harder to reason about.

The next evidence should be operational rather than promotional. OpenAI should publish measured completion rates for multi-day tasks, the frequency of human intervention, examples of permission failures and clear incident-reporting practices. Enterprises will also need auditable records showing what the agent read, decided and changed. Until those results exist, Dots should be understood as a significant new interface for delegated AI work—not as proof that autonomous digital labor is ready to run without sustained human oversight.

INNOVOX analysis

Dots matter because persistence changes the unit of AI work from an isolated prompt to an ongoing delegated responsibility. That could make agents materially more useful, but it also concentrates context, permissions and operational authority in a system that must decide when to act, pause and ask. The commercial winner may be determined less by benchmark intelligence than by reliability, auditability and permission design.

What to watch

Watch for measured task-completion rates, error and intervention frequency, enterprise audit logs, plugin-specific permission behavior and incident disclosure. Also track whether OpenAI expands availability beyond premium plans and whether users can inspect, export or delete the persistent notes and work history that make a Dot useful over time.