Most AI products wait at an empty box. OpenAI Dots are designed to do the opposite: remain present, keep work moving, and interrupt only when a result or decision needs you. That makes Dots one of the clearest signs yet that the personal-agent category has moved beyond chat memory and scheduled prompts.
The important question is not whether a dot can write an email or open a browser. Many agents can. It is whether OpenAI has assembled persistence, execution, app access and user control into something people can safely leave running. This guide separates what OpenAI has shipped from what it has merely previewed, then places Dots beside Today without pretending the two products are identical.
What OpenAI Dots Are
A dot is an always-on agent powered by GPT-6 Astra. OpenAI gives it a cloud computer, a browser and access to apps you choose to connect. It can work toward goals in the background, use other agents for parallel tasks, pause when it reaches a dependency and wake itself when work can continue. That description comes directly from OpenAI's launch announcement, published at DevDay on September 29, 2026.
You talk to the same dot through ChatGPT on desktop and web, with supported mobile access following desktop setup. Slack and Microsoft Teams can act as additional contact points, while texting is announced but not yet generally available. The promise is continuity: start a project in ChatGPT, add context in a team channel, and return without briefing a new assistant from scratch.
OpenAI says its plugin ecosystem gives Dots access to more than 4,000 apps. That number describes potential ecosystem reach, not blanket permission. A dot sees only the apps and scopes you connect, and each integration may expose a different mix of read and write actions. The Dots privacy and permissions FAQ says plugin permissions are shared across Dots, ChatGPT, ChatGPT Work and Codex, so connecting a service is a workspace-level trust decision rather than an isolated toggle for one agent.
The Cloud Computer Changes the Unit of Work
Ordinary chat turns are requests. A dot's unit of work is closer to an open loop.
It can investigate a bug while a team discusses it, prepare an invoice after noticing an omission, or keep a planning cycle moving as dependencies change. Its cloud browser keeps signed-in sessions separate from the browser on your laptop. Supported saved credentials can be used for sign-in without revealing the password to the model, and a new sign-in with a saved credential still requires confirmation.
Local access is a separate choice. You can connect one personal computer so the dot can work with local files and apps, but that computer must remain online with the ChatGPT app open. Granting that connection does not merge the cloud computer with your laptop; it expands the places the same agent can work. OpenAI's getting-started guide treats the dot's computer, connected personal computer and delegated Work or Codex environments as distinct execution surfaces.
This distinction matters. A cloud computer makes background execution possible. Local access makes the agent useful around files and software that never leave your machine. Neither one, by itself, proves the work will be correct.
Four Tests for Any Always-On Agent
1. Memory: what carries forward?
Dots learn from feedback and carry context across their channels. OpenAI also describes notes and proactive research that help a dot keep track of work. The current memory controls have a material limitation: OpenAI says users cannot view, delete or directly modify individual memories inside a dot's context. You can delete the entire dot context by deleting the dot, while files, conversations and ChatGPT memories created elsewhere remain separately stored. Disconnecting a plugin stops new access but does not erase information already incorporated into the dot's context.
That is a meaningful difference between persistence and governability. The product may remember enough to continue a project, yet still leave the user without a row-by-row account of what it believes. Before treating a dot as a long-term personal system, check what you can inspect, correct, delete and export—not only what the agent appears to remember.
If you are new to the category, our definition of an AI personal agent explains why durable, governable context matters more than a long chat history.
2. Proactivity: when does it speak first?
A dot can message you with progress, questions and decisions. It can pause and resume without requiring a fixed schedule for every follow-up. Its proactive research is deliberately narrower than ordinary execution: according to OpenAI, those background research tools are read-only and cannot send messages, change plugin content, or control a browser or computer. A follow-up action has to pass the normal action and approval rules. That separation reduces the chance that a passive scan quietly becomes an external action.
This is more adaptive than a reminder, but it creates a harder product problem: useful initiative can quickly become noise.
The quality test is not how often the agent contacts you. It is whether the intervention arrives while there is still time to act, explains why it surfaced, and lets you reduce or redirect future interruptions.
3. Execution: what can it finish?
The cloud computer, browser, plugins and background agents give Dots a broad execution surface. A dot can navigate websites, work with connected services, start cloud tasks and continue certain Codex work. OpenAI is also previewing specialist dots with their own organizational identity and credentials for well-defined company responsibilities.
Specialist dots should not be confused with the personal dot available in the initial rollout. Their deeper systems-of-record integrations, company-provisioned hardware and organizational identities are a preview direction, not evidence that every user already has a fleet of autonomous employees.
4. Control: where does the agent stop?
OpenAI gives users Custom Rules that allow, require approval for, or block particular actions. Built-in safety requirements still apply. Activity View shows background work, and automated review checks actions that could affect accounts or disclose information. Certain sensitive tasks, including password changes and money transfers, stay with the user; permanent deletion or software installation may require approval each time. An approval for one message does not become standing permission to contact that person again.
Those controls are not administrative decoration. They are the product. An always-on agent earns trust when the boundary between independent work, approval and human-only action remains visible under pressure.
Privacy and Safety: Concrete Protections, Residual Risk
The dot's cloud computer is isolated from your personal computer unless you connect the latter. OpenAI says website credentials can be used without exposing them to the model, and monitoring may pause or stop work when it detects a safety concern.
Data treatment depends on the plan. OpenAI states that Business, Enterprise and Edu workspace content is not used to improve models by default. Personal-plan users control model-improvement settings. The company also says it does not train directly on proactive research or a dot's private notes to itself, although information from those processes may inform an eligible conversation or task depending on settings.
None of this makes consequential work error-free. OpenAI's safety, security and privacy explanation treats prompt injection as an ongoing risk rather than a solved problem. A dot may misunderstand scope, follow a malicious instruction on a page or complete the wrong version of a task. Completed actions are not always reversible, even when the agent can help correct some document edits or recall an email.
The sensible adoption path is narrow access first: connect one low-risk source, keep external actions behind approval, inspect Activity View, then expand only after the agent behaves predictably.
How to Set Up a Dot Without Over-Permissioning It
OpenAI's setup flow starts in the ChatGPT desktop app or a desktop browser. The useful part is not giving the dot a name; it is defining a safe first responsibility.
1. Choose one bounded job. Start with a recurring research packet, project status review or document-monitoring task. Avoid mixing mail, payments, calendar changes and local files on day one. 2. Connect the minimum source set. Give read access only to the apps required for that job. Remember that existing plugin connections may already be shared across ChatGPT products. 3. Write explicit Custom Rules. State what the dot may draft, what it may change, who it may contact and which actions always need approval. Use names, destinations and limits rather than "handle this for me." 4. Keep external actions gated. Review proposed messages, purchases, deletions and software changes until you have evidence that the dot interprets your scope consistently. 5. Audit the first week. Check Activity View, correct misunderstandings as they happen, and disconnect sources that add risk without improving results.
The target is not maximum autonomy. It is the smallest permission set that produces a useful closed loop.
What Early Use Suggests—and What It Does Not Prove
The product is too new for reliable retention or task-success data. One detailed early-user report described a multi-day document review that maintained its place through interruptions, simultaneous work across the dot's cloud computer and a connected Linux machine, and a separate read-only background research agent. The same tester also found that delegated Work or Codex tasks use execution environments distinct from the dot's own computer, and that a failed launch did not automatically become a native queued task.
That one-day field report is useful because it exposes the execution model, not because it proves typical performance. It is one person's experiment with a new rollout, partly based on what the dot reported about itself. Treat it as a set of hypotheses to reproduce: sustained state, parallel environments, permission-aware continuation and unclear usage accounting.
The distinction also explains why opening several chats is not the same thing. Several chats can answer in parallel; a persistent agent can keep project state, identify blockers, delegate bounded work and return to interrupted tasks. Whether it does those things reliably over months remains unanswered.
What Is Available—and What Is Still Coming
As of October 6, 2026, OpenAI is rolling Dots out to Pro and Business Premium users in eligible markets. Enterprise users, including Edu and Healthcare workspaces, can enter the beta when an administrator enables it. Dots are not available to users under 18. The first personal dot is included with eligible plans, with an allowance for deeper work. Ordinary conversations with a dot do not count toward ChatGPT usage limits, but Work or Codex tasks launched by the dot use the corresponding allowances.
Several parts of the broader vision remain limited or planned:
- broader market and plan availability;
- texting as a contact channel;
- multiple personal dots and adjustable output capacity;
- specialist dots with enterprise identities and deeper integrations;
- Microsoft Agent 365 integration for specialist-dot governance.
Treat those as roadmap signals until they are available in the relevant plan, region and workspace.
OpenAI Dots vs. Today
Dots and Today point toward the same category: a personal agent that remembers enough context to help across days, notices when work needs attention, and can move tasks forward through tools. Grok Bot approaches the same shift as a roster of digital coworkers, while Manus Cue gives each agent a separate communication, payment and computing identity. Their centers of gravity differ.
| Decision | OpenAI Dots | Today |
|---|---|---|
| Starting point | OpenAI models, cloud computer and plugin ecosystem | Whole-day personal continuity |
| Persistent context | Cross-channel context and learning from feedback | Inspectable Living Memory designed around a person's ongoing life |
| Proactivity | Background work, progress questions and decision requests | Restrained proactive help tied to memory, time and authorized signals |
| Execution | Cloud browser, plugins, optional local computer and background agents | Connected-task execution with confirmation boundaries |
| Best initial fit | People already working deeply in ChatGPT and connected apps | People who want one personal agent to carry context across work and life |
This is not a feature-count verdict. Dots have formidable infrastructure and distribution. Today is making a different bet: that the durable advantage in personal AI will come from understandable memory, intervention quality and continuity across the whole day—not only the number of tools an agent can operate.
Today's product details should be judged from what is currently shipped, just as Dots should be judged from its current rollout rather than its specialist-agent roadmap. For the Today side, start with what Today is.
Who Should Try Dots First?
Dots make the most sense if your work already lives in ChatGPT, you have repeat projects that survive beyond one session, and you are willing to configure connections and approval rules. They are especially compelling for people who want research, coordination or software work to continue while they move between devices.
Wait—or start with a tightly limited trial—if your highest priority is fully inspectable personal memory, if most of your context lives in unsupported apps, or if you expect unsupervised automation to be correct by default. Always-on is an operating condition, not a guarantee of judgment.
The Real Shift Is From Sessions to Relationships
Dots matter because OpenAI is changing the shape of the product. The assistant is no longer a page you visit for answers; it is an ongoing identity with a computer, permissions, unfinished work and reasons to contact you.
That raises the bar for every personal agent. Memory must be understandable. Initiative must be worth the interruption. Execution must stop at visible boundaries. And the relationship has to become more useful over time without quietly becoming more invasive. Dots make that competition much more concrete—and give users a better set of questions to ask before they leave any AI running.