Tibo Sottiaux wants people to spend less time configuring AI and directing every step of its work. The head of ChatGPT and Codex at OpenAI describes an assistant that remembers a person’s goals, continues working between conversations and can be reached wherever it is needed.
Speaking with Lenny Rachitsky at OpenAI’s DevDay, he explains how the newly launched Dots platform fits that vision. He also discusses sharing revenue with plugin developers, the strain of working with multiple agents, and the skills he increasingly looks for when hiring.
He does not think the way people work with AI has settled. Today’s tools still feel awkward to him, and he expects substantial further changes.
An assistant that keeps working between conversations
For a comparison with Meta’s persistent agent, see the account of Muse.
Sottiaux describes Dots as an attempt to build an assistant that stays active, learns a user’s preferences and takes on ongoing responsibilities.
His longer-term vision is that the same assistant could join a meeting, take notes, continue the conversation through email and respond when called or messaged. People would not need to sit in front of a laptop directing every step.
The initial release starts with one primary dot. Sottiaux says OpenAI wants to learn how people use it before allowing them to create additional dots with specific roles. He already uses a separate one for work related to Twitter, which he describes as substantial enough to warrant its own agent.
OpenAI’s launch documentation distinguishes this first personal dot from specialist dots for organizations, which are beginning with focused enterprise pilots. The broader idea of building a team of personal dots remains a plan rather than something everyone can already do.
Sottiaux says the research behind the product goes back more than two years, including work on memory and tasks that continue over long periods. Dots brings those capabilities together with the infrastructure needed to keep an agent running.
Faster agents change how he works
Sottiaux still merges code occasionally, but says Codex now writes much of the code he needs for analysis, understanding the business and assessing product launches. Apart from occasional programming exercises at weekends, he no longer writes it by hand.
The number of agents he uses changes with their capabilities. When he pushes into a new kind of work, he often creates a larger team of agents. When a stronger model arrives, one agent can sometimes handle work that previously needed several, so he reduces the team again.
Speed has also changed the experience. Earlier, he missed the uninterrupted concentration of programming directly. Faster agents have helped him regain a version of that feeling, particularly when he can direct them by voice.
He expects manually configuring agent workflows to become less necessary. Rather than specifying exactly how an agent should repeat steps or move through a predetermined sequence, he wants a system that learns from the user’s goals and feedback.
Bringing ChatGPT’s capabilities together
Rachitsky asks whether Dots could become the main way people interact with OpenAI’s tools, with ChatGPT and Codex doing work behind it.
Sottiaux says the aim is an assistant available across different screens and communication channels. He also says OpenAI plans to bring Dots’ capabilities into ChatGPT and reduce the distinctions users currently have to navigate.
One immediate example is the separation between ChatGPT’s ordinary chat experience and its work mode. According to Sottiaux, users appreciate the work mode’s capabilities but sometimes prefer ordinary chat because it is faster and more pleasant. He says OpenAI is merging those experiences to reduce the need to choose.
Dots already avoids a model picker. Users talk to the assistant rather than selecting a model or setting how much reasoning it should perform.
Sottiaux admits that he finds these choices tiring himself. His goal is to remove as much configuration as possible.
Plugin developers could share in the revenue
Asked which DevDay announcement deserves more attention, Sottiaux points to OpenAI’s ecosystem for outside developers.
One part is Sign in with ChatGPT, which lets people use their subscription allowance in participating products. Another is opening more of ChatGPT’s own interface to plugins, so developers can build experiences inside it.
Sottiaux says OpenAI will share revenue with developers whose plugins receive substantial usage. Rachitsky uses Notion and Figma as examples: if subscribers use those services through ChatGPT, the companies could receive a share of the money. Sottiaux confirms that this is the intended arrangement, without giving payment rates or detailed terms.
He says plugin recommendations will depend on whether people keep using a plugin, its quality and whether it helps them accomplish something. A plugin can be recommended during conversations, but those recommendations can stop if it does not perform well.
His emphasis is on making something people continue to find useful, rather than simply getting a plugin listed.
Reducing the fatigue of working with agents
Rachitsky raises two concerns about AI-assisted work: constant switching between tasks and spending more time talking to agents than to other people.
Sottiaux says his team thinks about both. He wants to reduce configuration and make AI part of collaboration between humans, rather than leaving each person alone with a collection of agent conversations.
He imagines colleagues discussing an idea in a room, drawing something on paper and having an assistant begin building it in the background. They could put the result on a screen and continue developing it together through conversation.
He says the technology has not fully reached that point.
Rachitsky also asks about the pressure to do more simply because AI makes more possible. Sottiaux hopes assistants can reduce demands on people’s attention, help them attend fewer meetings and leave them better rested.
He compares that possibility with the different, more creative thinking he notices after disconnecting on holiday. He says the industry still has to work out how to make AI reduce that pressure in everyday life.
The dot that warned him before a failed demo
Sottiaux gives an example from DevDay. He says his dot alerted him that ChatGPT’s production service was down about five minutes before a live demonstration.
In his account, the agent connected the upcoming event with the service it depended on and recognized that he would want to know. It also offered to try fixing the problem.
He declined that offer because he did not think it was ready to do so, and contacted the engineering teams instead.
The agent had access to some production systems, but Sottiaux says that access came with guardrails. He describes specialist dots operating with additional monitoring and, in some cases, their own hardware, including Mac minis.
He also explains that the software coordinating a dot runs separately from the devices it uses. A dot can have its own computer and, with permission, connect to a user’s laptop or other devices. It is not confined to one machine.
What he looks for in people
Sottiaux says typing speed matters less as agents write more code. He increasingly values judgment about what makes a product good, understanding its users and caring about the audience it serves.
He sees the boundaries between engineering and design becoming less rigid. People who previously felt constrained by one role can now do more of the work involved in building a product.
Asked about early-career advice, he describes a colleague who joined OpenAI as a new graduate and quickly took on substantial responsibilities. What impressed him was the person’s kindness, willingness to collaborate, focus on important problems and ability to learn quickly.
Sottiaux says he has also revised his expectations about younger workers. He has been struck by how readily they adopt new tools and learn ways of working without being as attached to older habits.
For him, human creativity and judgment remain central to how the tools should be built. He also expects people to remain interested in what other people create, even as AI takes on more of the work.
How OpenAI develops products
Sottiaux describes a company where ideas often begin with small groups experimenting, rather than a fully specified project assigned from above.
His example is the Decisions API, a tool for choosing among predefined answers. He says it began with a small group working over a weekend. Other employees tried it, found additional uses and joined the effort.
That freedom comes with responsibility. Teams are trusted to make decisions, fix problems quickly and learn from mistakes. Sottiaux acknowledges that he caused early production outages himself, including one shortly after joining OpenAI.
He says the company still holds releases back when they need more work. Not every product considered for DevDay was ready to launch.
A future with more agent activity, and more monitoring
Sottiaux predicts that agents will eventually perform most actions on the internet. He expects models to become cheaper and faster, and text, voice and visual interaction to work together more smoothly.
He points to OpenAI’s work with Notion as an example of the demands that could create. Once agents could use Notion through MCP, a standard for connecting AI systems to tools, he says the service received substantial additional traffic. Products would need to handle that activity and work out how to pay for it.
He also wants more investment in experiences designed for humans, especially ones that make fuller use of voice and other forms of interaction.
On safety, Sottiaux describes spending more computing resources on systems that monitor agents while they work. These systems can check for risky actions and attempts to manipulate an agent through instructions hidden in the material it reads.
He says OpenAI has held back more capable models while continuing to improve safety and security. He is optimistic that the company can address these risks and describes a deep responsibility for what it puts in users’ hands.
The product change he most wants is simpler interaction. He wants people to spend less time choosing models and managing settings, and more time expressing what they want the assistant to do.