OpenAI's Dots Assistant Grew Out of Its Codex Coding Tools

Your daily brief on leadership and business

OpenAI has built its always-on personal assistant, dots, out of the agentic coding tools it first made for software engineers. Alexander Embiricos, who leads the company's enterprise product team, explained the path at TechCrunch Disrupt this month.

OpenAI released dots at its developer day in late September. It is available in the ChatGPT desktop and mobile apps for subscribers on the Pro tier or higher, with plans starting at $100 a month. Once connected to a user's calendar, email, computer and other tools, it tries to help get work done even when the user is not asking it anything.

Embiricos said much of that could already be done with the Codex coding app, released at the start of this year, but it took more work and hand-holding. His team spent its effort on how the product is framed. Presenting it as a single entity in a single conversation, he said, "starts to become intuitive". The name is a play on the negative space inside OpenAI's logo and is meant "to feel like just something that could eventually be ubiquitous".

Proactive by design

The point of a persistent agent is that it needs no prompting. For OpenAI that means more token usage. For users it means an assistant that is thinking about how to help. In the reporter's testing, the ability to think ahead was apparent and often useful, but not quite there yet. It flagged calendar conflicts and deadlines, though not always in good time. Embiricos recalled that his own dots once warned him of a conflict three minutes before it started.

He also described an engineer at a drone delivery startup whose dot had noticed an outage, analysed it and come to him with information on how to resolve it before the engineer knew anything was wrong.

Data and controls

Embiricos said dots includes a feature called proactive research. On a regular cadence it reviews what it knows about the user, what the conversation has covered and what it has been given access to, then researches how it could be more helpful and proposes it. OpenAI says the agent has the same privacy controls as ChatGPT. Users can turn off model training and control which resources it can reach.

TechCrunch found those controls are not granular, and that the app refuses to accept read-only access to cloud drives. Embiricos said OpenAI needs to build "super, super, super granular controls". He added that the company launched with conservative defaults, so the agent asked permission often, and drew criticism for it.

Meta's assistant, Muse, has been criticised for the profiles it builds of users' family and friends. Hark, a start-up building its own assistant, says it does not collect data to fuel an advertising business.

On whether personal agents can scale, Embiricos noted that he keeps a human assistant because OpenAI's corporate partners want to talk to a person. He said he never intends to have a meeting with someone else's agent. OpenAI is in conversations with payment processors, he said.

Source: TechCrunch