OpenAI Dots: Persistent Agents, Limited Access, and Constraints

OpenAI unveils Dots, always-active AI agents that run on their own cloud machine and integrate with over 4,000 applications. Access remains restricted, and gray areas persist regarding privacy, compliance, and reliability, as GPT-6 Astra is presented as an advancement for multi-step tasks.

Restricted Access, Regional Exclusions, and Compliance Gray Areas

Dots is initially offered to ChatGPT Pro subscribers starting at $100 per month, as well as Business Premium users. Free offerings, Go and Plus, do not include this service. Pro subscribers located in the European Economic Area, Switzerland, and the United Kingdom are excluded at this stage. OpenAI has not communicated specific compliance terms, dedicated availability guarantees, or pricing for additional agents beyond the first. For teams subject to security and procurement requirements, these gaps weigh heavily in the evaluation.

Non-Editable Memory and Persistence of Plugin Traces

A Dot learns from feedback and retains context, but users cannot view, modify, or delete individual memories. Disconnecting a plugin does not erase the context already memorized from it. OpenAI clarifies that Dots can make mistakes and that impactful work should be reviewed. Dots are not presented as full automation.

What Dots Technically Changes Compared to Previous Agents

Dots correspond to persistent agents operating on their own cloud computer, capable of remaining active after the user's machine is turned off and connecting to over 4,000 applications. Each agent runs GPT-6 Astra, pursues a permanent goal between conversations, and does not require a new prompt at each step. Access is via ChatGPT, Slack, Microsoft Teams, or voice, but at launch, agents do not have a standalone email address, cannot initiate calls, and SMS sending is limited to a Pro beta in the United States. Compared to previous attempts, often fragile beyond three or four steps, GPT-6 Astra is presented as more reliable for multi-step and multi-tool sequences. A Dot retains context, learns a model of preferences, and operates on persistent goals, distinguishing it from a feature limited to a session.

Usage Examples and Delegation Principle

Among the demonstrations, one agent identifies an invoice to send, extracts information from an email thread, and then drafts the document for approval. Another scenario shows monitoring customer feedback, preparing fixes with tests, and then creating pull requests with videos for review. Sam Altman describes the relationship as a delegation to an agent already familiar with the context, learning preferences and retaining corrections. Unlike the one-off exchange of a chatbot that stops at the end of the session, these agents maintain a work thread.

For Data Teams: Targeted Tasks, Permissions, and Control

The challenge is more about responsibility than raw power. Where a chatbot requires step-by-step guidance, a persistent agent calls for clear objectives, appropriate permissions, and integrated checkpoints in the process. For data scientists, the potential lies in tracking statuses, monitoring results, updating documentation, and summarizing long conversations, which could shift from hours to minutes with good memory and application access. The risk is a relaxation of supervision: on a pipeline or model evaluation, tight permissions and explicit stop conditions are necessary, especially since proactive behavior and persistence require setting limits before activation.

What Remains to Be Proven and the Role of User Framing

The ability of Dots to handle changing exploratory analyses, long searches, or collaborative environments remains to be demonstrated, and the examples presented are chosen to showcase the product. The real test will focus on ambiguous and large-scale tasks. Although persistent agents have been announced multiple times without fully delivering, this version is presented as the most credible, while shifting the requirement to the user: clarifying what is delegable, what requires approval, and which data should not pass through a hosted agent. Many practitioners have not mapped their work at this level of detail, as the agent is only as good as the brief provided, and this matters more than ever now. Dots marks a shift from reactive to proactive use, but adoption requires informed skepticism.