Alexander Embiricos breaks down ChatGPT Dots: proactive personal agents with their own identity, compute, and Slack presence

ChatGPT Dots: OpenAI's Proactive Personal Agents

Alexander Embiricos, a product manager at OpenAI, has spent two years waiting for this product. The chat paradigm was informational. Coding agents like Codex added execution but still required constant direction. Dots, launched at DevDay, are the step Embiricos says makes the difference: proactive agents that notice what needs doing and act before you ask.

What a dot actually is

Each dot has its own cloud computer, its own Slack identity formatted as [username]-dot, and access to anything its owner can do on a machine. It can also connect to a local setup if the user has one. The naming convention matters: earlier experiments with random agent names like "Wisp" created confusion about who was doing what. The owner-linked format makes the delegation relationship legible to everyone in the channel.

The proactive framing is the sharpest departure from prior agent products. Embiricos describes a scenario where someone urgently needs a headshot: rather than waiting to be asked, a dot notices the request, locates the file, and surfaces it with a confirmation prompt. The user's involvement shifts from initiating tasks to approving them.

“Dot, it's like you have that principal engineer because your Dot has its own computer. It can do anything that you can do on a computer in the cloud. It can also connect to your computer... They show up with like your username, so like I'm AE. It's like AE dash dot in Slack. And so you have all these people and their delegates just doing stuff together.”

Dots versus specialist dots

OpenAI is drawing a deliberate line between personal dots and specialist dots. Personal dots extend the individual, using that person's connectors and acting under their identity. Specialist dots are IT-managed, centrally governed, and deployed by organizations to accelerate specific functions across a team. Procurement and accounting are the early use cases where Embiricos says OpenAI has seen the most traction. OpenAI also announced a partnership with Microsoft around identity governance for specialist dots, though Embiricos doesn't detail the terms.

The identity design choice

A previous internal approach had agents sending messages under a "sent by ChatGPT" label. It worked, but Embiricos says it created noise. The current design draws a hard line: when you collaborate with your dot on an email and hit send, it looks like it came from you. When the dot acts independently, it has its own distinct identity in Slack. That separation, he argues, is what keeps communication legible at scale, particularly when multiple people and their delegates are all active in the same channels.

Who benefits most

Embiricos singles out solopreneurs and small business owners as the audience he is most excited about. Engineers already have Codex. The small business owner drowning in inbound tasks is the person who gains the most from an agent that identifies what needs doing without being asked. He also flags video game developers as a category gaining fast, given that dots run on Astra, the same model underlying recent generative game-building demos.

Plugin extensions and UI

OpenAI is giving third-party developers a direct path to ChatGPT users through plugin extensions: custom UI and pages that render inside the ChatGPT app. A meetings plugin launched at DevDay is the first example. Embiricos is skeptical of fully on-the-fly UI generation for regular use, arguing users need some predictability to build mastery, but sees plugin extensions as the right middle ground between generated UI and locked-down native surfaces.

The adoption bottleneck

The biggest constraint on diffusion outside San Francisco, Embiricos says, is that most people don't know to ask the product what it can do. He treats that as a product failure, not a marketing one. Dots are partly designed to close that gap: if users don't know to ask for help finding a file, a dot that proactively surfaces the file removes the need to ask in the first place.

Dots run on GPT-4o Astra. The unlock, in Embiricos's framing, isn't a smarter model. It's removing the requirement that users know what to ask for.