OpenAI launches dots, bringing ongoing work into ChatGPT
Bottom linePersistent agents shift the product’s promise toward follow-through. The practical test is whether useful autonomy can coexist with clear, manageable permission boundaries.
What changed
OpenAI introduced dots on September 29, describing agents that can continue working between conversations. Powered by GPT-6 Astra, each dot has a cloud computer and can use connected applications. The company says users can communicate through ChatGPT, Slack and Teams, while the agent retains context across those channels.
The launch separates ongoing activity from unrestricted action. OpenAI says proactive research uses read-only connected tools. Users can inspect progress, choose application access and define action rules, with built-in safety requirements still applying. The company also previewed specialist dots for organizational responsibilities.
OpenAI’s release notes describe a gradual rollout to eligible adult Pro and Business Premium users. Pro access initially excludes the European Economic Area, Switzerland and the United Kingdom. Enterprise access is a beta that administrators must enable; the announcement is not universal availability.
FUVISIGHT analysis
The important product change is the unit of delegation. A conversation usually ends with an answer or an artifact. An ongoing assignment can instead require the assistant to notice new evidence, revise earlier work and decide when a person needs to intervene. That could reduce the cost of repeatedly reconstructing context, especially for projects scattered across several applications.
It also creates a harder usability problem. A user needs to understand both what the agent can see and what it may change. Those are separate decisions. A simple permission screen is useful only if the assistant’s behavior remains legible days later, after the original instruction has faded from memory.
For an organization, a sensible evaluation would start with a bounded workflow whose outcome can be checked: maintaining a project brief, preparing a recurring analysis or identifying changes that need review. Measure useful completed work, correction time and unnecessary interruptions. Counting messages or tool calls would reward activity without establishing value.
The strongest version of this product would earn broader responsibility through reliable small outcomes. The weaker version would create a second inbox of suggestions and approvals. Whether persistent assistance saves attention will depend as much on its restraint and reporting as on the model’s ability to complete an isolated task.
What to watch
Watch for measured completion rates in ordinary workflows, understandable incident histories and evidence that users can revise or withdraw permissions without losing track of ongoing work. The meaningful adoption signal is sustained delegation with manageable oversight, rather than the number of agents created.