OpenAI's dots give agents their own identity — the real news isn't autonomy, it's the permissions layer

OpenAI has launched dots, personal AI agents that take on continuous work at varying levels of autonomy, according to the company's presentation reported by Infobae. Each dot runs on GPT-6 Astra — a model OpenAI positions for research, coding and multi-step tasks — and comes with its own cloud computer, its own browser and access to connected applications. Sam Altman, OpenAI's CEO, said the agents can work around the clock, organize tasks, analyze information, write code and coordinate actions under rules, permissions and controls defined by each person or company. Rollout is gradual: Pro, Business Premium and Enterprise plans, with the first dot included at no extra cost in the first two, a quota for deeper tasks, expanded limits for the first month, and — notably — conversations with a dot not counting against normal ChatGPT usage limits. For now, access is limited to a small group of users.

The headline feature is "it works 24 hours." The consequential one is buried three paragraphs down: according to OpenAI, a dot has its own identity to administer permissions. That is a quiet but structural change. Until now, an AI agent acted as a puppet wearing your credentials — everything it touched was logged as you. An agent with a distinct identity can be granted scoped access, audited separately and revoked without locking out the human. It is, in practice, the first plumbing needed to treat software as a headcount line rather than a tool. Every enterprise security team that has spent a decade managing service accounts just acquired a much harder version of the same problem.

The second thing to read closely is the distribution surface. Dots connect to more than 4,000 applications through a plugin ecosystem, and you can talk to them from ChatGPT, SMS, Slack, Microsoft Teams or a voice call, with context carried across all of them. That is not a feature list — it is a land grab. We have argued here before that the frontier-model war has shifted from "who has the smartest model" to "who controls the plumbing": distribution, integration and agent standards. Shipping an agent that lives natively inside Slack and Teams is OpenAI declining to be a guest in Microsoft's and Google's houses. The model is the engine; the 4,000 connectors are the road network.

Now the discipline our editorial line demands: nothing here has been measured in public. There is no reliability figure, no task-success rate, no error budget, no cost per completed task. "Works continuously" is a description of uptime, not of competence — an agent that runs all night and is wrong at 8 a.m. has consumed compute, not delivered work. The gradual rollout to a limited group, the quotas on deeper tasks and the beta gating behind a workspace administrator all point to a system OpenAI itself does not yet consider settled. Note also which demos were chosen: a transcript turned into clips and show notes, customer feedback turned into scoped fixes and pull requests with tests, a scientific analysis refreshed with new data. All three share a property — the output is immediately checkable by a human. That is a tell about where the technology actually works today.

On jobs, the demos are the analysis. Reviewing transcripts, drafting social posts, triaging recurring complaints, scoping a fix, regenerating charts: that is the coordination and back-office layer, exactly the band we have flagged as most exposed across sectors. The work that survives is the work in the loop OpenAI describes — the agent sends updates, asks questions and requests approval when a decision requires it. The human moves from doing to judging. The failure mode is obvious and worth naming now: approval fatigue. A reviewer who clicks yes forty times a day stops being a control and becomes a rubber stamp, and the same identity-and-permissions machinery that makes dots useful is what an attacker would want most.

Our reading: this is a real step toward the long-term case we keep making — one person commanding a small team of tireless collaborators is how abundance actually arrives, and how more people end up working on what they care about instead of what their inbox demands. But the short term is a transition, not a gift. The scarce skill for the next few years is not prompting; it is reviewing well, and building organizations that audit non-human identities as seriously as human ones. Judge dots in six months on completed-task reliability and on how many approvals a user can meaningfully give per day — not on how many hours they can run.