OpenAI introduced Dots in its September 29, 2026 DevDay update: always-on agents that can continue working between conversations. A user gives a dot a goal, chooses which apps it can connect to, and defines what it may do on its own. The dot runs on its own cloud computer and browser, then brings results or decisions that need judgment back to the user.
That is different from a normal chat window or a one-off agent task. A dot takes on an ongoing responsibility, such as tracking project updates, preparing meeting briefs, following a code migration, or comparing new information with existing documents. It can continue after the user leaves the conversation and reach out through supported contact methods instead of waiting for the next prompt at every step.
OpenAI positions Dots between ChatGPT, connected apps, and Codex or Work workflows. A dot can use relevant conversation context and preferences, create files, handle data, and delegate parts of a task to ChatGPT Work or Codex. That makes it closer to a long-lived coordinator, but it also makes permissions, context accuracy, and state persistence more important than in ordinary question answering.
Controls are central to the product. Users can choose which apps to connect, manage permissions, define Custom Rules, and decide which actions require approval or must be blocked. Actions that could affect accounts or share information go through checks, and some still require human approval; high-impact steps such as changing a password remain with the person. This layered approval model matters as much as model capability for making persistent agents usable.
Dots are rolling out gradually, and access depends on plan, market, age, and workspace settings. Enterprise use requires administrator enablement, with some access still in beta. The workflows shown in the announcement therefore do not mean every user has the same capability today, or that every app connection exposes the same depth of control.
From a workflow perspective, Dots move agents from waiting for instructions to continuously watching a goal and reporting when needed. The benefit is that a long-running project does not have to be re-explained every time. The cost is that users must design permissions, stop conditions, data boundaries, and approval points more carefully. Once an agent can use a browser and tools on a cloud computer, a mistake is no longer just an inaccurate sentence; it can be a wrong document edit, an unintended disclosure, or a process that should have stopped earlier.
The meaningful shift is not simply that AI can work by itself. Dots put persistence, tool use, and responsibility boundaries into one product model. Early users should treat them as bounded collaborators, start with low-risk tasks that can be reviewed, retain an action trail, and let the agent stop when its context or permissions are insufficient. Persistent work should not be confused with unrestricted autonomy.
