OpenAI unveiled dots, a persistent assistant that runs continuously in the background, at its developer day in late September. The assistant is available inside the ChatGPT desktop and mobile apps to subscribers on the Pro tier or higher, with plans starting at $100 per month. Once connected to a user’s calendar, email, computer and other tools, dots aims to proactively help with work and scheduling tasks.
From chatbot to coder tool to always-on agent
Alexander Embiricos, who leads OpenAI’s enterprise product team, explained at TechCrunch Disrupt how the product evolved. He described a convergent trend among companies building assistants: treating the agent as a single persistent entity and one continuous conversation rather than many separate chats makes the interaction feel more intuitive. OpenAI’s prior work — including the Codex coding tool released earlier this year and the spinout of ChatGPT Work from developer-oriented agent tools — paved the way for repackaging model capabilities into a more accessible, always-on assistant.
Embiricos noted that much of dots’ functionality could be achieved with Codex, but that dots requires less manual setup and is intended to be easier for nontechnical users to adopt.
Proactivity and current limits
A key selling point for dots is proactivity: the agent will surface calendar conflicts, upcoming deadlines and other issues before a user asks. In practice, this feature is still improving. Embiricos recounted an early example where dots notified him of a calendar conflict only three minutes before the event, which left little time to act. He expects timing to get better with iteration.
He also shared a concrete example of practical utility: an engineer at a drone delivery startup reportedly found that their dots had already detected, analyzed and prepared useful remediation suggestions for an outage before the human operator noticed it.
Privacy, permissions and control
As with other personal agents, dots raises questions about how much personal data an assistant needs to be useful. Embiricos described a feature called “proactive research,” where, on a cadence, the agent reviews what it knows about the user, what conversations it has had, and what resources it can access, then proposes ways it could be more helpful.
OpenAI says dots inherits the same privacy controls as ChatGPT: users can disable model training so OpenAI does not use their conversations to improve its models, and they can control what resources the agent may access on their device. However, those controls are not highly granular at present, and the app can be insistent about permissions — for example, it currently rejects read-only access to cloud drives.
Embiricos acknowledged the need for much more granular permissioning and emphasized the model’s intelligence and discretion as part of the trade-offs. He said the product launched with conservative defaults that requested permission frequently, and they expected user feedback that they plan to tune over time.
Can agents scale?
A broader question is whether personal AI agents can scale beyond early adopters. Agents can be compelling in a few hands — stories of snagging reservations or automating tasks are common — but businesses and people often prefer to interact with humans. Embiricos said he uses a human assistant because OpenAI’s corporate partners want to deal with people; he also said he personally does not intend to attend meetings held with someone else’s agent.
OpenAI is cautious about rolling out agent abilities that act in the external world and stresses coordination with the rest of the ecosystem. Embiricos said the company is in deep conversations with payment processors, suggesting that allowing an assistant to transact on a user’s behalf is a possible, though complex, next step. Enabling payments would be lucrative but also a significant ask for consumers who may be uncomfortable granting that level of access.
Conclusion
Dots represents OpenAI’s attempt to move from developer-focused tools toward an always-on, proactive assistant for everyday users. While it demonstrates useful behaviors and potential real-world benefits, its success will depend on improving timeliness, expanding granular privacy controls, and resolving practical scaling and partnership challenges before such agents become widely adopted.
