Tibo Sottiaux: Why AI Agents Will Take Most Actions on the Internet

Tibo Sottiaux, who leads ChatGPT and Codex at OpenAI, told Lenny’s Podcast at DevDay that most actions on the internet will soon be taken by AI agents. He sat down with Lenny Rachitsky in front of a live audience, hours after his team shipped more than 20 products, including Dots, OpenAI’s new always-on personal agent. The 37 minute conversation covers where agentic AI is heading, how ChatGPT plugins will make money, what skills still matter, and how OpenAI thinks about AI safety.

TLDW

Sottiaux argues that hand-built agent loops and workflow graphs are a passing phase, and that the future is one persistent agent that knows your goals, learns from feedback and shows up on any screen. Dots is OpenAI’s first version of that, and Codex, ChatGPT and Dots are being merged so users stop choosing between products and models. He calls the open ecosystem the sleeper hit of DevDay: Sign in with ChatGPT, plugins ranked by retention, and revenue sharing with developers. He closes on what builders are not pricing in (agents as the main users of the internet, much cheaper and faster models, seamless multimodality) and on why OpenAI is spending heavily on monitoring and holding back its most capable model.

Thoughts

The most useful idea in the first few minutes is the one Sottiaux tosses off almost as an aside. He says his own teams of agents grow as he pushes at the edge of what a model can do, then collapse back to one agent each time a better model arrives. That is a clean way to think about every piece of agent orchestration being sold right now. Loops, graphs, routers and role prompts are compensation for a model that cannot yet hold the whole job in its head. They have value, but it is the value of scaffolding, and scaffolding comes down. Anyone whose product is mostly orchestration should ask what is left when the next model makes the team of ten into a team of one. His own answer is that what stays is memory, preferences and feedback, which is where Dots is aimed.

The ecosystem section deserves more attention than the Dots demo, and Sottiaux says so himself. Plugins are recommended inside conversations based on retention and quality, not on keywords or who wrote about you. Developers whose plugins get used are paid a share of revenue, and partners who adopt Sign in with ChatGPT let users spend their existing ChatGPT usage inside another product. That is an app store whose ranking signal is whether people keep coming back. It also means OpenAI decides what gets surfaced to 1.2 billion users, and a plugin that slips in quality simply stops being recommended. Lenny notes this is the second attempt at a marketplace. The difference this time is that there is money on the table and a ranking system that is hard to game with marketing. Whether developers trust a single gatekeeper with their distribution is the open question, and the interview does not press on it.

The best story in the episode is a small one. Five minutes before the DevDay live demo, Sottiaux’s Dot messaged him to say ChatGPT production was down, then offered to try to fix it. He declined and called the engineers. Read that carefully, because it shows where agents really are. The impressive part was not an action. It was relevance: the agent connected an outage, a keynote, a demo that depended on the failing system and a five minute window, and decided a person needed to know. The part he refused was the action. That is a sensible line for anyone deploying agents today. Let them watch everything and interrupt you well, and be slow to hand over the keys. It also explains the architecture he describes next, where the agent’s harness does not live on your machine and specialist Dots run with extra guardrails on their own hardware.

The headline claim comes late, around the 31 minute mark, and the interesting part is the second-order effect. When Notion shipped an MCP server, agents started doing real work through it and traffic arrived in volume. That traffic strains systems and does not behave like human traffic, so the economics have to be worked out again. Pricing per seat assumes a person in the seat. Rate limits assume human pace. Advertising assumes eyes. Sottiaux says a company can hold back from building an agent interface for a while, but that it is inevitable. I think he is right, and the uncomfortable follow-on is that products will be chosen by agents on reliability and cost, not on how they look. He pairs this with a point that is easy to miss: delightful experiences built for humans are under-invested in. If agents do the routine work, the human-facing part of a product has to earn its place.

On safety, the concrete detail is better than the reassurance. Sottiaux describes a growing share of compute going to secondary monitoring, a second system watching the working agent for high-risk actions and prompt injection, and says OpenAI has not released its next step up in capability. That is a specific, checkable posture. The argument he leans on at the end is weaker. Saying that a company serving 1.2 billion people cannot afford to get it wrong is an incentive, not evidence, and incentives cut both ways when rivals are shipping. The more honest note in the conversation comes earlier, when he says the promise of AI is not one more prompt per second and that he thinks more clearly after a week fully disconnected. An industry that sells always-on agents while its leaders do their best thinking offline has not yet resolved that tension, and he admits as much.

Key Takeaways

Sottiaux no longer writes code by hand at work. Codex writes the code he uses for analysis of trends, the business and how launches are doing. His only hand coding is the occasional LeetCode problem on a weekend, which he calls therapeutic.
His agent teams expand and shrink. He builds bigger teams of agents as he pushes the frontier, then a model breakthrough lets one larger agent do the whole job and keep it in memory, and the team shrinks again.
He thinks fiddling with loops is not how agentic work will end up. The goal is a system that learns what you want to achieve, so you do not have to design the loop yourself.
Dots launched with one primary Dot per user, available first to Pro users. Multiple Dots with specific roles are coming soon. Sottiaux already runs a second one whose whole job is monitoring Twitter for him.
Codex and ChatGPT are being merged, including the chat and work toggle, and the capabilities of Dots will eventually ship straight into ChatGPT for what he says are 1.2 billion users.
Dots has no model picker and no configuration beyond choosing which channels to talk to it on. Sottiaux says he gets tired of model pickers and reasoning settings himself.
Sign in with ChatGPT launched with 16 partners. It began last year as an informal handshake with the creators of Pi and OpenCode, who were allowed to use the Codex sign-on on trust.
Popular plugins will receive a share of revenue, a detail that was not in the keynote. Plugins are recommended in conversations based on retention and quality, and stop being recommended if they are not good.
The research behind Dots is more than two years old: long-horizon persistent tasks and memory systems, combined with the ability to run around the clock. In the video he says Dots launched on Astra, which he calls OpenAI’s safest and most aligned model.
A Dot’s harness does not run on your machine. It has its own computer and can connect to many devices at once, which he compares to an octopus. Some specialist Dots inside OpenAI run on Mac minis with added guardrails and monitoring.
The skill trending down is typing fast. The skills trending up are taste, thinking about the user and connecting with the audience you build for. He says OpenAI employs more than 120 former Y Combinator founders.
Three things he says builders are not pricing in: most actions on the internet will be taken by agents, models will get cheaper and faster at remarkable rates, and all modalities will finally work together seamlessly. His test is to imagine everything ten times better in a year and ask whether you would build differently.
He changed his mind on three things in the past year. Model capability arrived a year or two sooner than he expected, he relies on voice far more than he thought he would, and younger people have turned out to be the first to embrace and harness the change, which altered his hiring strategy.
On safety, he describes pacing the frontier as investing ahead in alignment, security and guardrails, with more and more compute spent on a second system that monitors the working agent. He says OpenAI has not released the next step up in capability beyond its current top model.

Chapters

01:12 How Tibo Works With Agents, and Why Loops Are a Passing Phase

Sottiaux opens by saying his team worked through the night in a library turned war room while he slept. He merges code now and then, but Codex writes nearly all of it. He used to run many agents in parallel, and a faster model has put him back in a state of flow with fewer of them. His teams of agents grow and shrink with each model generation. He does not think people will keep tuning loops by hand, and expects a system that learns your goals to replace that work.

04:13 Permanent Active Intelligence and One Merged ChatGPT

The vision is an always-available intelligence that joins your meeting, takes notes, picks the thread up over email and answers a text, then gets out of the way. He says carrying a laptop everywhere ties you to the technology when it should be working for you. Work will keep changing radically, and today’s tools still feel clunky. Codex and ChatGPT are merging, and Dots capabilities will land in ChatGPT itself. Asked about the film Her, he says he thinks more about older science fiction such as Neuromancer and the original Star Trek, where you talk to a computer and it acts.

11:34 The Sleeper Hit: An Open Ecosystem and Plugin Revenue Sharing

Asked what from DevDay is underrated, Sottiaux picks the ecosystem. Sign in with ChatGPT has 16 partners and grew out of a handshake deal with the makers of Pi and OpenCode. OpenAI is opening its plugin infrastructure and plugin discovery so developers can reach its full user base. Plugins that see real usage will share in revenue. His advice for getting discovered is to build a good plugin, because recommendations follow retention and quality.

14:39 The Research Behind Dots

Lenny asks why a personal agent took so long, given OpenClaw and the rival assistants that followed. Sottiaux jokes that the Codex Cloud animation from a year ago inspired a competitor’s bot. His serious answer is that long-horizon tasks and memory have each been in research for more than two years, with much of the memory work already shipped in ChatGPT. Dots combines those with a harness that runs around the clock. He says safety and security took a large share of the effort.

17:25 Agent Fatigue, Loneliness and the Pressure to Do More

Lenny raises context switching and the loneliness of talking to agents all day. Sottiaux says his team thinks about this constantly, starting with less configuration and making agent work less of a solo activity. His ideal is an agent present in the room that listens to a conversation between people and builds in the background. On the pressure to run 30 agents at once, he says the promise is less noise and attention spent where you want it. He notes that he thinks more creatively after a week fully disconnected.

20:09 The Dot That Flagged an Outage Before the Live Demo

Five minutes before the DevDay demo, Sottiaux’s Dot told him ChatGPT production was down and offered to try a fix. He turned the offer down and contacted the engineering teams, who resolved it. What struck him was that the agent linked the event, the production system and the timing without being asked. His Dot has access to some production systems behind guardrails. Specialist Dots add monitoring and run on their own hardware, and a single Dot can control many connected devices.

23:25 Skills Trending Up and Down, and Advice for New Grads

Typing fast is out. Taste, user focus and knowing what good looks like are in, which is why former founders do well and why Lenny thinks product managers will thrive. Roles are blurring, so a designer or engineer who felt boxed in now has room. Faster models and voice control have given Sottiaux back a creative flow he missed. For people early in their careers he points to Ahmed Ibrahim, hired as a new graduate and now responsible for OpenAI’s applied compute fleet, who stands out for kindness, collaboration and learning at unusual speed.

27:18 How OpenAI Ships: Bottoms-Up Energy, Autonomy and Mistakes

Sottiaux describes a company full of former founders working from the bottom up. One new API began as four people hacking on a weekend in a Slack channel, then spread as others got excited and pitched in. Leaders hold a quality bar, and several launches were held back from DevDay to ship in the following weeks. People get a lot of autonomy and are expected to own the results. He admits he took production down on his third day at OpenAI and kept his job.

30:45 Building for an Internet Where Agents Do Most of the Work

This is the core claim. Agents will take the majority of actions online, models will get far cheaper and faster, and modalities will merge. Most products he sees are not built with that in mind. Notion’s MCP server brought a surge of agent traffic, which strains systems and forces new economics. He also says human experiences that use every modality are under-invested in. He then lists what he got wrong: the pace of capability, the importance of voice, and how quickly younger people would adapt.

33:55 AI Safety, Pacing the Frontier and the End of Model Pickers

Sottiaux defines pacing the frontier as investing ahead in alignment, security and guardrails. A growing amount of compute goes to a secondary system that watches the primary agent and steps in on risky actions or prompt injection. OpenAI has released a more efficient model near its top level of intelligence, and has held back the next step up. He says he is proud of that restraint. His closing complaint is about his own product: the app should nearly disappear, and nobody should need to understand model pickers and reasoning settings.

Notable Quotes

“Having to set up and fiddle with your loops is something that maybe people got excited about, but I don’t think this is the way that it’s going to work.”
Tibo Sottiaux, on hand-tuned agent workflows

“It still feels like a bit clunky, and I think it will feel clunky until it isn’t.”
Tibo Sottiaux, on working with AI today

“We may not have coders anymore, but we have more builders than ever.”
Tibo Sottiaux, on where people stay valuable

“I took production down like day three.”
Tibo Sottiaux, on his first week at OpenAI

“I think the majority of actions on the internet will be taken by agents.”
Tibo Sottiaux, on what builders are not pricing in

“We build products for 1.2 billion people. We can’t screw that up.”
Tibo Sottiaux, on OpenAI’s incentive to get safety right

“You kind of need a PhD in model pickers.”
Tibo Sottiaux, on what annoys him about the ChatGPT app

Model names, product details and user numbers here are as stated in the interview, recorded at OpenAI DevDay and published on October 4, 2026. Watch the full conversation here.