OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux

Summary · 4 min read instead of 37 min

Tibo Sottiaux leads ChatGPT and Codex at OpenAI, where his teams have shipped Codex, ChatGPT Work and the Dots personal agent platform. He says he was an engineer for 10 years. He spoke at DevDay, hours after his team launched more than 20 products.

Sottiaux argues that setting up loops and graphs for agents is a passing phase, and that users will instead want one smart agent that runs 24-7, learns their goals and preferences, and is reachable from any device. He says Dots, plus a planned merger of Codex and ChatGPT capabilities, is OpenAI's step toward that, with no model picker to configure. He calls the open plugin ecosystem the sleeper launch: sign in with ChatGPT has 16 partners, popular plugins will get a revenue share, and recommendations are driven by retention. He expects most internet actions to be taken by agents and urges builders to imagine models about 10 times better within a year. On safety, he says OpenAI spends more compute on secondary monitoring and held back a model it had trained.

Minute by minute

1. 1:34

   Codex writes his code now. Sottiaux says he no longer writes code by hand, except for occasional weekend practice. Codex writes the code he uses to analyze trends, the business and launches. He says the size of his agent teams grows as he pushes the frontier, then shrinks when a model breakthrough lets one bigger agent do everything, and that a faster ultra fast model put him back in flow.

2. 3:08

   Hand-built agent loops will fade. Asked about the loops and graphs trend, Sottiaux says fiddling with loops excited some people but is not how things will work. He wants an agent that works 24-7, understands goals and preferences and learns from feedback, which is how Dots is positioned. He imagines it joining a meeting, following up by email or text, so people stop carrying a laptop around like a brick.

3. 6:14

   Work will keep changing radically. Sottiaux expects work to keep changing, saying voice and multimodal tools are coming together but still feel clunky. He describes a shared whiteboard where humans and agents collaborate, beyond today's chat clients. OpenAI will merge chat and the work toggle and eventually ship Dots capabilities into ChatGPT for its 1.2 billion users, and he likes that Dots has no model picker.

4. 10:20

   Users will run teams of dots. Each user currently gets one primary dot that learns their preferences most deeply. Sottiaux says users will soon be able to add more dots to form a virtual team with specific roles. He has one dedicated to monitoring Twitter, a job he says is the work of an entire dot.

5. 11:31

   The open ecosystem is the sleeper. Sottiaux calls the open ecosystem the under-noticed launch. Sign in with ChatGPT, which began as a handshake with the creators of Pi and Open Code, now has 16 partners. Popular plugins will get a share of revenue, and plugins are recommended in conversations based on retention and quality, so weak ones stop being recommended.

6. 14:34

   Dots rests on years of research. Sottiaux says OpenAI has worked on long-horizon tasks for more than two years and on coherent memory systems for about as long. Dots combines that memory with the Codex harness and the ability to run 24-7. He says the team put extra effort into safety and launched on Astra, which he calls OpenAI's safest, most aligned model.

7. 17:17

   AI should reduce noise, not add pressure. Rachitsky raises loneliness, context switching and pressure to run more agents. Sottiaux says the team works on reducing configuration fatigue and imagines agents present in physical space during human conversations. He says he thinks more creatively after a week offline, and that AI's promise is fewer distractions and possibly fewer meetings, not more prompts.

8. 20:22

   A dot flagged a DevDay outage. Sottiaux's dot pinged him five minutes before a live demo that ChatGPT production was down, linking the outage to the demo, and offered to fix it. He declined and brought in engineering teams. He explains that the Dots harness does not run on the user's machine; it can connect to many devices, and specialist dots run with extra guardrails and monitoring, some on Mac minis.

9. 23:22

   Taste beats typing fast. Sottiaux says typing fast matters less and taste, user focus and knowing what good looks like matter more. He says OpenAI has more than 120 former founders and that roles are blurring. For early-career people he points to Ahmed Ibrahim, hired as a new grad and now responsible for OpenAI's compute fleet and Applied, citing his kindness, collaboration and speed of learning.

10. 27:06

    Bottoms-up shipping with high autonomy. Sottiaux describes the Decisions API starting as four people hacking on a weekend, spreading through internal use and shipping. OpenAI holds a quality bar and spaced out some DevDay launches. He says people get autonomy and must own mistakes; he took production down on his third day and was still employed afterward.

11. 30:44

    Most internet actions will be agents. Sottiaux says builders are not pricing in that agents will take most internet actions, that models will get cheaper and faster, and that modalities will merge. He urges imagining things roughly 10 times better in a year. Notion's MCP drew heavy agent traffic, he says, so products must be built for scale and new economics, while human experiences remain underinvested.

12. 32:38

    Capabilities and voice arrived sooner. Sottiaux says he expected today's model capabilities a year or two later and had to revise his priorities. He did not expect to rely so much on voice, dictation and calling his agent. He also changed his view on hiring after seeing how younger generations embrace the change first and learn fast.

13. 34:06

    Pacing the frontier means investing ahead. Sottiaux says OpenAI spends more compute on secondary monitoring of agents to catch high-risk actions and prompt injection, and that most API stack investment goes to safety. He says OpenAI did not release 6.1 Astra and is proud of that. He wants to remove the model picker and reasoning settings, saying users should not need a PhD in model pickers.

Published Oct 4, 2026.