OpenAI introduces Dots: from chats to agents that keep working

OpenAI has introduced Dots, agents designed to retain work context and continue tasks through connected tools. The proposed change goes beyond a written answer: the system takes part in a sequence of work. [OpenAI]

In a chat, the conversation usually moves forward when the user writes again. A persistent agent is organised around an objective that may require several actions. This difference explains the industry's interest, and why discussing agents also means discussing permissions.

A working environment, not just a model

The 29 September announcement describes a cloud computer for each agent, with a browser and app connections. OpenAI also proposes proactive information gathering and research, while retaining action controls. The announced rollout includes pilots with business specialists; it does not establish simultaneous general access for every account. [OpenAI]

Consider a hypothetical example: summarising several folders and drafting an email. Knowing which files the agent consulted helps a reviewer check the summary; seeing the recipient and content before sending allows a separate action to be authorised. If those stages blur together, it becomes hard to understand what was delegated. A persistent agent makes this visibility more important, since work may continue after the initial conversation.

What can it do, and with whose permission?

Looking up information, preparing a document and sending it have different consequences. A tool involved in those steps needs limits its users can understand. They must be able to distinguish a prepared proposal from an action already taken.

This matters particularly when connecting applications. Access that is useful for reading documents may be excessive if it also allows edits. The system's value should be assessed alongside the clarity of its controls.

From a message to a verifiable result

An agent demonstration can look convincing because it shows many consecutive steps. To evaluate it, examine the result: does the file exist, is the recipient correct and are the underlying data appropriate? The appearance of work is not enough.

Our reading is that this stage shifts some attention from the model to its surrounding environment. Connections, limits and the ability to review completed actions become part of the product.

What still needs checking

Rollout and access conditions can change. Usage guides and plan comparisons need current information. For an initial news report, the priority is to explain the proposal and distinguish announced capabilities from independently tested ones.