OpenAI has launched Dots, AI agents designed to take responsibility for ongoing work instead of waiting for a new prompt at every step. Announced at DevDay on 29 September, Dots can keep projects moving between conversations, coordinate tasks and return with results or decisions that need human input.
Each Dot runs on GPT-6 Astra and has its own Cloud computer and browser. OpenAI says its plugin ecosystem provides connections to more than 4,000 apps. The promise is an assistant that follows work as it changes, rather than another separate chat to manage.
There are important qualifications. Access depends on subscription and region, proactive research has read-only restrictions, and a Dot’s retained memory cannot currently be edited item by item. Those details matter as much as the launch demonstrations.
What can OpenAI Dots actually do?
OpenAI positions Dots around continuing responsibilities. Its examples include a developer’s agent investigating customer feedback and preparing tested improvements, a sales agent checking requirements against product documentation, and a creator’s agent turning interview material into show notes and draft social posts.
The common thread is follow-through. Instead of producing one document and stopping, a Dot is intended to revise work when requirements change, remember feedback and identify the next useful step. OpenAI’s examples describe intended workflows, not independent evidence that every stage will work reliably.
Users can reach the same Dot through ChatGPT, Slack and Microsoft Teams. Relevant context can carry across those channels, although conversations remain separate and access to private information does not automatically authorise sharing it with colleagues.
Dots can also delegate work to parallel agents. That makes the product a way to coordinate tasks, not simply a new name for the underlying model. Our coverage of GPT-6 Astra’s computer-use capabilities explains the model behind that approach.
Who can use Dots at launch?
The rollout is gradual, so an eligible subscription does not guarantee immediate access. OpenAI’s Help Centre sets out an important difference between personal and business availability.
| Plan | Launch availability |
|---|---|
| ChatGPT Pro | Rolling out in eligible markets, excluding the UK, the European Economic Area and Switzerland. |
| ChatGPT Business Premium | Available across supported ChatGPT regions, subject to rollout. |
| ChatGPT Enterprise, Edu and Healthcare | Beta access when enabled by a workspace administrator. It is off by default. |
For UK readers, the distinction is practical: the initial exclusion applies to Pro, rather than every business deployment. Business Premium is the specifically named business tier.
Initial setup requires the ChatGPT desktop app or a desktop browser. Mobile app conversations become available with the supporting rollout, but creating a Dot on mobile and using it through mobile web are not supported at launch.
An included Dot does not mean unlimited task execution
According to OpenAI’s Dots launch announcement, the first Dot is included with Pro and Business Premium at no extra cost. Conversations with it do not count towards ChatGPT usage limits.
The Help Centre also says Dot usage will not count towards eligible plans’ allowances for the next month, with subsequent usage terms to be announced. However, the launch announcement separately states that tasks a Dot starts or manages in Codex or ChatGPT Work count towards those products’ usual limits.
The distinction is between access to the agent and the work it commissions. An always-available assistant should not be read as a promise of unlimited coding, analysis or document production.
Cloud work can continue when your computer is off
A Dot’s Cloud computer is separate from the user’s own machine. OpenAI’s documentation says Cloud work can continue while personal devices are switched off, and users can inspect the agent’s computer or take over when necessary.
Connecting a personal computer is optional. Work that depends on that connection requires the machine to remain online with the ChatGPT app open. Cloud availability therefore does not make an offline laptop’s files and applications remotely usable.
A similar distinction applies to integrations. Connecting a messaging channel is not the same as granting access to an inbox, repository or local computer. Nor does adding a Dot to Slack automatically create a monitoring task: users must specify what it should watch and establish a supported trigger or schedule.
Proactive research is not permission to change everything
OpenAI distinguishes assigned work from what it calls proactive research. The latter lets a Dot examine permitted connected information and retain private notes about potentially useful findings without waiting for a specific request.
Those research tools cannot directly send messages to other people, alter connected app content or control a browser or computer. Follow-up actions must pass the normal permissions and safety checks. Separately authorised recurring tasks can include actions, so not all background activity is read-only.
Custom Rules let users add boundaries, such as prohibiting outgoing emails. A separate Auto-review system checks certain proposed actions against instructions, rules and safety requirements. Some sensitive steps, including changing passwords and transferring money between financial accounts, must be handed back to the user.
OpenAI says monitoring can pause work when it detects a safety concern, but acknowledges that Dots can still make mistakes. These safeguards are controls to evaluate, not proof that an agent cannot send the wrong information or misunderstand a task.
Persistent memory comes with a significant limitation
A Dot can retain information from conversations and connected services. However, OpenAI says users currently cannot view, delete, or directly modify individual Dot memories. Disconnecting an app stops new access through that connection; it does not erase information already retained.
Deleting the Dot removes its own context, but separately stored files, Codex threads and ChatGPT conversations remain. Information shared with ChatGPT Memory is managed separately, too.
For Business, Enterprise and Edu workspaces, OpenAI does not use content for model training by default. Personal-plan users control eligible training use through the “Improve the model for everyone” setting. Proactive research is not used directly for training, but information brought from it into an eligible conversation may be, depending on settings.
For teams, that makes deciding what to connect an important first step. Disconnecting a service later doesn’t mean the agent forgets everything it learned from it.
The useful test is whether Dots reduces supervision
For creators and small businesses, the most useful starting point is a bounded responsibility: prepare a weekly research brief, maintain a draft launch checklist, or turn approved source material into content for review. Publishing, customer commitments and production changes can remain outside that initial scope.
The test should measure more than whether a result appears. Did the Dot use the latest instructions? Did it respect the agreed boundary? How much correction was needed? Did it notice a failed action rather than report success?
OpenAI is separately piloting specialist Dots with their own organisational identities and credentials, and working with Microsoft on Agent 365 integration. Those pilots are distinct from the initial personal Dot rollout.
DIY AI has not independently tested Dots’ reliability at launch. Its potential value is clear: less time repeatedly supplying context and directing the next step. Whether it delivers that value will depend on completing work correctly and the supervision required to get there.
