Introducing dots: OpenAI's Always-On Agents With Their Own Cloud Computers

On September 29, 2026, OpenAI introduced dots, a new kind of AI product. OpenAI describes dots as "remarkably capable, always-on agents built to handle everything." A dot is not a chat window that waits for your next prompt. It is a long-lived assistant that learns your goals and standards, works in the cloud on its own, and brings finished work back for your review. OpenAI calls it "a whole new way to work with AI." Start with the core facts from the announcement. Dots are powered by GPT-6 Astra. Each dot has its own cloud computer and its own browser. Through OpenAI's plugin ecosystem, a dot can connect to more than 4,000 apps. It learns from feedback over time and can work toward your goals around the clock. You reach your dot through ChatGPT, Slack, or Teams. You can ask questions, explore ideas, give feedback, or start a voice call. Dots are starting to roll out across the Pro, Business Premium, and Enterprise plans in eligible markets, and OpenAI says it plans to expand to more users soon. The announcement page gives no pricing, no benchmark scores, and no latency figures. This article does not invent any. The mechanism has three layers worth separating. The first layer is the execution environment. A dot runs on its own cloud computer with its own browser and the apps you have connected. You can open your dot's computer at any time and inspect its work. This matters. It turns an agent acting inside a black box into an agent whose workspace a human can look at. A dot can also connect to other devices. If you give permission, it can use your laptop and work directly beside you. The second layer is persistence and parallelism. OpenAI says a dot can take a project and run with it while it works on several others. You can keep bringing it new tasks without juggling separate threads or directing every step. The third layer is memory and learning. The more you work with a dot, the more it learns your preferences, how you think, and what good looks like to you. OpenAI calls this the magic of dots: they bring you work done the way you would do it, sometimes before you think to ask. From an engineering view, the product center of gravity moves away from single-turn prompt answers. It moves toward long-running task state, a model of personal preference, and event-triggered action.

The scenarios in the announcement are specific. A developer's dot watches customer feedback for recurring requests, scopes smaller improvements and bug fixes, builds and tests them, and returns complete pull requests with videos that show the changes. A product lead's dot learns the audience, positioning, and creative standards. When scope changes, it reworks the story, revises launch materials, and drafts asset and document changes. A scientist's dot reruns analyses when new data arrives, investigates unexpected results, updates the figures for a paper, and flags what needs human review. A sales lead's dot checks customer requirements against product docs, finds what still needs testing, builds a proof of concept for a key integration, and suggests a solutions engineer who handled a similar deal. A content creator's dot finds clip moments in a new interview transcript, prepares show notes, and drafts social posts for approval. Inside OpenAI, the team describes a bug appearing in Slack and dots starting to investigate at once. A new design arrives and dots turn it into a working app. A planning cycle starts and dots keep everyone in sync toward the deadline. Outside OpenAI, an early tester's dot noticed he had forgotten to invoice a publication, prepared the invoice, and sent it after his approval. That last example shows the basic rhythm: the dot notices, the dot prepares, and the human approves. Safety and control have their own section on the page. It has three parts: built-in safeguards, access and permissions, and action review and approvals. The stated promise is that you are always in control. The invoice example shows one concrete behavior: a consequential outward action waits for human approval. The public page does not describe the technical detail behind these safeguards. It does not cover prompt-injection defenses, permission granularity, or audit-log formats. For enterprises, these are exactly the items to verify before any rollout.

The organizational piece is a preview of specialist dots. A specialist dot has its own identity for access management, runs on IT-provisioned hardware, and supports deep integrations with a company's systems of record. That lets it take on well-defined responsibilities inside the organization. OpenAI also says it is bringing specialist dots to Microsoft Agent 365. This step matters. An agent stops being a personal tool and becomes a digital worker in the identity directory, with an account, a permission boundary, and a way for IT to manage and revoke it. There are four practical implications for developers and enterprises. First, the entry point of work changes. Before, a person opened a tool, wrote a prompt, and checked the output. Now the dot watches event sources, such as a bug in Slack, customer feedback, new data, or a new transcript, and prepares results unprompted. Second, review becomes the main labor. When a dot delivers a full pull request and a demo video, human value sits in judgment and approval. Teams must redesign review flows and accountability. Third, identity and permission governance becomes a launch gate. Every dot needs its own credentials, an explicit list of reachable apps, and least-privilege limits on actions. Fourth, the plugin ecosystem becomes a moat. Reach into more than 4,000 apps means integration breadth is itself a product feature.

The challenges are equally clear. Reliability comes first. An always-on agent can turn a small error into a chain of errors, so action review and rollback matter more than any model score. Cost is the second challenge. OpenAI has not published pricing or how compute is metered, and a cloud computer that runs around the clock is an unknown line in an enterprise budget. Data governance is third. A dot learns personal preferences and keeps access to many systems, so retention, isolation between users, and compliance rules must be explicit. Accountability is fourth. When a dot has done most of the preparation before you approve, approval can shrink to a quick click. Review fatigue becomes a real risk. For the industry, dots move competition from "whose model is stronger" to "whose agent has better identity, permissions, memory, and integrations." OpenAI says you can start today with a primary dot, give it a name, and make it your own, and that it envisions teams of dots working together on your behalf over time. The tie to Microsoft Agent 365 shows that the management layer for agents is merging with the enterprise identity stack. Signals to watch in the short term include the speed of expansion to more markets and plans, the release date of specialist dots, the pricing model, and independent third-party assessments of the safeguards. The advice for readers is practical. Developers should start with work that repeats, can be verified, and has clear acceptance criteria, such as bug triage and small fixes. Enterprise IT should define a dot's identity, permission scope, and approval rules before any pilot. In every role, treat inspectability as a hard requirement. If you can open the dot's computer and see its actions, trust has something to stand on.