Build for the Agent Majority

Tibo Sottiaux leads ChatGPT and Codex at OpenAI. Under him, OpenAI has shipped Codex, ChatGPT Work and Dots, its new always-on personal agent platform. A longtime engineer, he now has Codex write nearly all of his code and keeps hand-coding as a weekend hobby. He sat down for this conversation in front of a live audience at OpenAI's DevDay, hours after his team launched more than 20 products. These signals are from Tibo Sottiaux's interview with Lenny Rachitsky on Lenny's Podcast.

1. The 10x Assumption. Sottiaux looks at what most people are building and thinks they're "not quite getting it." He names 3 things the market hasn't priced in: most actions on the internet will be taken by agents, models will get cheaper and faster at startling rates, and every modality will finally work together. His advice is to assume everything is roughly 10 times better within a year and build for that world. Products designed for today's models will look dated almost immediately.
2. Agents as the Main User. When Notion shipped its MCP server, agents suddenly started doing work through it, and traffic surged. That load strains systems and forces companies to work out the economics of serving machines. Sottiaux says companies can delay building an agent interface for a while, but the shift is inevitable. He also thinks new experiences for humans, built around voice and other modalities, are underinvested.
3. Learning Over Configuration. Setting up and tuning agent loops and graphs is something "people got excited about," Sottiaux says, and he doesn't think it lasts. He wants one smart agent that runs 24/7, knows your goals and preferences, and learns from feedback. Dots has no model picker; the only setting is which channels you use to reach it. His biggest annoyance with ChatGPT today is the menu of models, reasoning efforts and modes: "you kind of need a PhD in model pickers."
4. Off the Laptop. Sottiaux wants an intelligence that follows you across any screen: it joins the meeting, takes notes, picks the thread back up in email and answers your texts. He calls carrying a laptop everywhere being tied to the technology. Over the past year he changed his mind about voice. He now does a lot of his work by dictating or calling his agent, which he hadn't expected.
5. Agent Teams Grow and Shrink. When he pushes on hard problems, Sottiaux builds larger and larger teams of agents. Then a model improvement arrives, one bigger agent can do the whole job, and the team shrinks again. Faster models put him back in a creative flow state he'd lost as a manager. Dots launched with one primary agent per user, and multiple dots with specific roles are coming; Sottiaux already runs one that monitors Twitter for him.
6. The 5-Minute Warning. About 5 minutes before his live DevDay demo, Sottiaux's dot pinged him that ChatGPT production was down. It had connected 3 facts on its own: DevDay was happening, the demo probably used that production system, and he'd want to know. It offered to fix the outage, and he politely declined. Specialist dots touching production run with extra guardrails and monitoring on their own hardware, some on Mac minis, and the agent sits off-device so it can control many machines, "a little bit like an octopus."
7. Distribution for Builders. Sottiaux calls the open platform the sleeper hit of DevDay. Sign in with ChatGPT now has 16 partners; it started as a virtual handshake with the creators of Pi and OpenCode. Plugin makers such as Notion or Figma will get a revenue share when ChatGPT subscribers use their usage inside those products. His advice for getting discovered: "Build a good plugin." OpenAI recommends plugins in conversations based on retention and quality, and stops recommending ones that underperform.
8. Memory Takes Years. Dots rests on more than 2 years of research into long-horizon tasks and roughly as long on memory. Sottiaux says many users have a story like this one: someone with a mystery burn on their hand had ChatGPT recall an earlier barbecue with limes and tequila, and it identified a lime-induced burn. A lot of the extra launch effort went into safety and security, which is why Dots launched on Astra, which Sottiaux calls OpenAI's safest and most aligned model.
9. Taste Over Typing. The skill falling in value is typing fast. Rising skills are taste, thinking about the user and connecting with the audience you build for. OpenAI now has more than 120 former YC founders, which Sottiaux says makes it a "mega startup." Role boundaries are blurring, so people who sat uneasily between design and engineering now have an advantage.
10. The Next Tibo. Asked for advice to new grads, Sottiaux points to Ahmed Ibrahim, hired at OpenAI as a new grad and now responsible for the company's compute fleet and applied work. Ibrahim stood out for being kind and collaborative, solving the most important problem without putting himself first, and learning faster than anyone Sottiaux had seen. Sottiaux now trusts him with OpenAI's gnarliest launches. He also revised his view on hiring: younger people adopt the new tools first and absorb them fastest.
11. Autonomy With Ownership. Many OpenAI launches start bottom-up. A new API began as 4 people hacking on a weekend, became a Slack channel, drew in volunteers and shipped. Leaders then hold the quality bar, and several DevDay launches were held back and spaced into the following weeks. Sottiaux took production down on his third day and kept his job. He says he could have pushed teams harder toward building less complex things.
12. Pacing Through Safety. Sottiaux defines pacing the frontier as investing in alignment, security and guardrails ahead of capability. OpenAI spends a growing share of compute on secondary monitoring that watches the primary agent for risky actions or prompt injection. It has not yet released anything beyond Astra-level capability, and he's proud that it held back one model. With 1.2 billion users, many of them nontechnical, "we can't screw that up."

Watch the full video at https://www.youtube.com/watch?v=MM-C3JqCXBk.