OpenAI Dots: Always-On ChatGPT Agents With Cloud Computers
OpenAI wants ChatGPT to stop waiting for the next prompt. At its DevDay developer conference on September 29, the company introduced Dots, a set of always-on AI agents inside ChatGPT. Each one runs on its own cloud computer and can keep working on goals without instructions at every step.
The idea is simple: instead of trading messages one request at a time, a user gives an agent a goal and lets it carry the work forward. It can juggle several projects at once, learn its owner's preferences from feedback and, according to OpenAI, sometimes deliver work before anyone has asked for it.
How Dots works
The agents run on GPT-6 Astra, OpenAI's current top model. Each agent has a dedicated cloud computer, and users can open it at any time to see what the agent is doing. That visibility matters for a system meant to act on its own.
Through OpenAI's plugins, the agents can reach more than 4,000 apps. With permission, an agent can also work directly on the owner's laptop, not just in its cloud machine.
Users can talk to their agent in several places:
- ChatGPT on desktop, web or mobile, by message or by voice call
- Slack
- Microsoft Teams
Context travels with the agent across these channels. A project that starts in ChatGPT can move to a team in Slack without anyone briefing the agent again. Along the way, agents send progress updates and ask their owners questions.
Early use inside and outside OpenAI
OpenAI says its own teams have been using the agents. When someone reports a bug in the company's Slack, agents start investigating right away. New designs are turned into working apps, which leaves staff free to focus on customer feedback.
One outside tester offers a smaller example. He had forgotten to bill a publication. His agent noticed, drafted the invoice and held it until he approved sending it.
Proactive research, with limits
When an agent has no assigned task, it looks for ways to help. OpenAI calls this proactive research. During that mode, the agent's connections to the owner's apps are read-only. It can look, but it cannot send messages or change anything.
Other safeguards sit around the agent rather than inside it. Any action that could touch a user's accounts or share information must pass a check called auto-review. A separate monitoring system can pause or stop an agent if it raises a safety concern. OpenAI is also direct about the limits: it tells users to review consequential work because the agents "can still make mistakes."
Specialist agents for companies
OpenAI also previewed specialist agents. An employer sets them up with their own identity and credentials for one defined job. Inside OpenAI, the tested jobs included procurement and invoice processing.
Customers begin with focused pilots, where OpenAI engineers help define what each agent is responsible for. Microsoft is working with OpenAI so these agents can be managed through the governance and security controls in Agent 365, Microsoft's product for overseeing agents in the enterprise.
Competition from Meta
Dots puts OpenAI in direct competition with Meta Platforms, which launched its Muse personal agent in the U.S. on September 8. Muse also runs on a dedicated cloud computer. Meta says most uses are free, with subscriptions for heavier users.
Availability and pricing
Dots began rolling out on September 29 to ChatGPT Pro and Business Premium subscribers, though some accounts may wait several days. Pro subscribers in the European Economic Area (the EU plus Iceland, Liechtenstein and Norway), Switzerland and the U.K. are excluded for now. Enterprise customers, including those on Edu and Healthcare plans, can try a beta once a workspace administrator enables it.
One agent is included with a Pro or Business Premium plan at no extra cost. OpenAI's help documentation says agent usage will not count against plan allowances for the first month; plan-specific terms follow after that. Extra agents, and more speed or monthly capacity for existing ones, are planned for later. OpenAI has not disclosed prices.
Our Take
Dots is less about a new model and more about a new operating pattern: agents that persist, hold context across tools and act without being prompted. That shift raises the stakes. An agent that answers questions can mislead you; an agent with 4,000 app connections, laptop access and a Slack presence can act on that mistake.
OpenAI appears to understand this. The design puts several controls outside the model: read-only access during proactive research, an auto-review gate for sensitive actions, a monitor that can halt an agent and a cloud computer users can inspect. This suggests the company treats model behavior and system permissions as separate layers of defense, which is the right instinct. Whether auto-review holds up against prompt injection arriving through connected apps is still an open question, and it is worth watching how independent testers probe it.
The pricing and rollout also say something. Starting with premium tiers, excluding Europe and deferring prices points to a cautious, capacity-limited launch. It also fits OpenAI's broader push on cost, seen in its work on cheaper models approaching Astra. If always-on agents are to reach more users, running them cheaply will matter as much as capability.
For readers, the choice is becoming clearer. Cloud-hosted agents from OpenAI and Meta offer convenience and reach. Others may prefer local coding agents on their own GPU, trading capability for control. Watch next for the post-trial pricing, the European launch timeline and how enterprises govern specialist agents with their own credentials.