From a chat request to an ongoing responsibility
OpenAI used its 29 September DevDay update to introduce dots, describing them as always-on agents that can take responsibility for work rather than waiting for each new prompt. A user gives a dot a goal, defines what it may do on its own and can continue talking to it while the work proceeds. The dot is intended to return with results or ask for judgement when a decision cannot be delegated. Rollout is gradual to eligible accounts, so the announcement should not be read as immediate universal availability.
The distinction from a conventional assistant is continuity. A one-off chat usually ends when the answer is delivered. A persistent agent may need to monitor progress, revisit evidence, use connected applications and keep a task moving across days. That makes the quality of its memory and the limits of its authority as important as the quality of the model producing each response.
A computer and a wider set of tools
OpenAI’s documentation says dots are powered by GPT-6 Astra and live in the cloud with their own computer and browser. They can work while the user’s computer is off, drawing on relevant previous conversations and preferences. The examples include research, data analysis, document preparation and software development. Setup can offer application connections and, in the desktop app, access to the user’s computer. Each connection changes what the dot can see or do and should be reviewed as a deliberate grant of access.
The company says users create a dot from the ChatGPT desktop app or a desktop browser, then can speak to it as it learns their work. A dot can be given a name and visual appearance, but the operational issue is the identity behind that interface. Teams need to know which connected account an action uses, which files it can reach and whether a task continues when its owner is away. Personalisation may make the agent easier to work with; it does not replace permission design.
Conversations across channels
OpenAI describes a dot that can be reached in ChatGPT, Slack, Teams or a call, subject to available channels and account access. Switching channels does not create a new dot or reset its memory. That is useful when a decision starts on a laptop and continues on a phone, but it also complicates information boundaries. Messages stay in their respective channels, while the dot may use relevant context across them. The documentation says it checks permission before sharing information from a private conversation with others.
Users can tell a dot where to send different updates, such as routine progress in ChatGPT and decisions requiring attention elsewhere. OpenAI says adding a dot to a Slack channel alone does not begin monitoring: the user has to specify what to watch. This distinction matters. A team should be able to tell whether the agent is a passive participant, an active monitor or a delegated actor with permission to respond. Those roles should not be blurred by a friendly interface.
The hard part is deciding what can run unattended
An always-on agent may discover a changed deadline, a failed build or an unanswered request when no one is watching. It can be valuable if it prepares options and surfaces the right decision at the right time. It can be disruptive if it floods channels, makes assumptions about priorities or acts before the context is complete. A sensible starting task has a narrow objective, known data sources, a clear review point and a rule for when to stop and ask.
Persistent memory also needs care. A dot may draw on earlier preferences, but preferences can become stale and private context may not belong in a shared workspace. Organisations should set expectations about retention, access, auditing and offboarding before placing an agent in operational channels. The same safeguards that govern human account access should apply to the connected systems an agent operates.
What the launch does and does not establish
DevDay establishes the product direction and the 29 September announcement date. The detailed documentation describes current behaviour and availability caveats; it does not prove that every advertised workflow will be reliable for every user. Early trials should measure whether the dot notices relevant changes, produces verifiable work, asks for a decision at appropriate points and respects its assigned scope. Human review remains particularly important where actions are irreversible or affect other people.
Dots point towards an assistant relationship built around continuing responsibility. Their usefulness will be determined by the unglamorous mechanics of permissions, memory, notifications and recovery as much as by model capability. For Australian teams considering the rollout, a small supervised pilot is a better test than immediately connecting a broad set of business systems. Pilot owners should document when the agent is expected to pause, what it may do independently and who reviews its outputs.