In 30 seconds
Tibo Sottiaux argues that fiddling with loops and agent graphs is a passing phase, and that people will soon rely on a persistent agent, Dots, that learns their goals and works across devices. He also covers OpenAI's open plugin ecosystem with revenue sharing, the skills that matter in an agent-heavy workplace, the culture inside OpenAI, and why he thinks most internet actions will soon be taken by agents.
- 1Hand-tuning loops and graphs is a passing phaseSottiaux says people ultimately want a system that learns what they want to achieve, instead of one they have to configure exactly.
- 2The goal is technology that gets out of the wayA permanently active intelligence should be available through any client or screen, and recede when not needed.
- 3The "sleeper hit" is the open ecosystemChatGPT sign-in partners, plugins and revenue sharing let outside developers reach 1.2 billion users, with discovery driven by retention.
- 4A Dot warned its owner of a production outage five minutes before a live demoIt connected the dots between DevDay, a production system and the demo, and offered to fix it.
- 5Taste and knowing what good looks like are rising in value; typing fast is notOpenAI is full of former founders, and Sottiaux thinks product managers will thrive.
- 6Most actions on the internet will soon be taken by agentsProducts must plan for agent-scale traffic, and builders should imagine everything being roughly ten times better in a year.
Read the full transcript by chapter
Agent teams expand and shrink with each model breakthrough, and hand-built loops are a passing phase
Sottiaux, who leads ChatGPT and Codex at OpenAI, says he rarely writes code by hand any more. Codex writes the code behind his analyses of trends and launch performance, and it also writes much of the code for Dots. His only hand-coding is the occasional weekend LeetCode problem as a therapeutic moment.
He used to run many agents in parallel, but a recent breakthrough in model speed put him back in the flow with fewer. He describes a cycle: as he pushes the frontier he builds larger teams of agents, then a new model breakthrough lets one bigger agent do it all, with memory, and he shrinks the team again.
Asked about loops and graphs, he says that setting up and fiddling with loops may have excited people but is not how it will work. He wants an agent that understands goals and preferences and learns from feedback, so users do not need to specify how to loop it to get good results.
Dots aims to be a permanent, always-available intelligence with no model picker
Sottiaux describes the broader goal as breaking free from technology through a permanent, active intelligence that is available on any client or screen. It could appear in a meeting room to take notes, be picked up over email or answer a text. The aim is for it to be there when needed and out of the way otherwise, instead of tying people to a laptop.
He expects work to keep changing radically, since voice, multimodal inputs and outputs are only now coming together and still feel clunky. He imagines talking to an agent that remembers everything and working on a shared whiteboard where humans and agents collaborate.
OpenAI plans to merge Codex and ChatGPT's chat and work modes, and eventually bring Dots' capabilities into ChatGPT to lift the floor for its 1.2 billion users. Dots has no model picker and nothing to configure except which channels to talk through. He cites Neuromancer and Star Trek as influences, where you could simply talk to a computer or a spaceship.
The open ecosystem, with plugin revenue sharing and retention-based discovery, is the quiet big bet
At launch each user gets one primary Dot, which learns preferences most deeply and can be texted. Sottiaux says users will soon be able to add more, building a virtual team with specific roles. He already has one assigned to Twitter, a job large enough to occupy a whole Dot.
When Lenny asks what is being overlooked, he names the ecosystem. Sign-in with ChatGPT started informally last year with creators of tools like Pi and OpenCode and now has 16 partners. Plugins and extensions can reach ChatGPT's users, and popular plugins will receive a share of revenue, as will products where people sign in with ChatGPT and use their subscription usage.
For discovery, the advice is to build a good plugin. OpenAI looks at retention and quality and recommends plugins to users in conversations, and low-quality ones stop being recommended. It is about consistent real usage rather than optimizing keywords to get people writing about you.
Dots rests on years of memory and long-horizon work, with safety built in, and humans stay at the center
Sottiaux traces the lineage of Dots to Codex Cloud about a year ago, and says the research on persistent, long-horizon tasks and memory systems has been under way for more than two years. Much of the memory work already shipped in ChatGPT, where people share stories of it remembering personal details. Dots combines this with the Codex harness and the ability to run around the clock.
Safety and security took a lot of additional effort, and Dots launched with OpenAI's safest and most aligned model, which he calls Astra. On where humans stay valuable, he says OpenAI designs products to put humans at the center as extensions of your will and taste. There may be fewer coders but more builders, and humans will still want to hear from and learn about what other humans build.
On context switching, loneliness and pressure to do more, he says the team thinks constantly about reducing configuration fatigue and the solo-adventure feel of talking to agents. His ideal is an agent in your physical space that hears ideas and builds in the background. He hopes AI will reduce noise and let people focus, not demand one more prompt per second.
A Dot flagged a production outage before the live demo, and architecture and guardrails make that possible
Sottiaux's example of a striking AI moment is that his Dot messaged him five minutes before the DevDay live demo to say production was down. It reasoned that the event, the production system and the demo were connected, and offered to try to fix it. He told it, "I don't think you're there yet", and contacted the engineering teams, who fixed it.
On access, he says Dots can reach some production systems with guardrails, and specialist dots operate with extra guardrails, monitoring and their own hardware. Some run on Mac minis. The key design choice is that the harness does not run on the user's machine, so a Dot has its own computer and can connect to many devices, which he likens to an octopus.
Taste and ownership are rising skills, and early-career people can learn fastest with the new tools
Asked about skills, Sottiaux says typing fast is trending down, while good taste, thinking about the user and connecting with the audience are trending up. OpenAI has many former founders, including more than 120 he counts, and a passion to build something that matters and knowing what good looks like matter more than ever. Lenny adds that PMs should thrive, and Sottiaux says roles are blurring, which is a chance to shine for designers and engineers who felt uncomfortable in narrow lanes.
He lost the engineering flow state for a while but is getting it back with very fast models and voice control. For new graduates, he points to a colleague, Ahmed Ibrahim, hired as a new grad and now responsible for OpenAI's compute fleet and applied work. What makes him stand out, Sottiaux says, is kindness, collaboration, a focus on the problems that matter and the ability to learn quickly with the latest tools.
OpenAI moves through bottoms-up experiments, high autonomy and a high quality bar
Inside OpenAI, Sottiaux says, many ideas come bottoms-up from ex-founders. He describes a decisions API that began with a few people hacking on a weekend using a strong model and constrained sampling, then grew as a Slack channel attracted more colleagues, who added things like visual inputs. The process feels unstructured, and people assume roles to channel it productively.
The company holds a high quality bar, and some things planned for DevDay were held back to space out launches over the following weeks. He says people get a lot of autonomy and must own the results, and when something goes wrong they fix it quickly and learn. That includes him, since he took production down on his third day at OpenAI and kept his job.
Agents will take most internet actions, so build for agent scale and make safety investment pace the frontier
Sottiaux says the things not yet priced in are that most internet actions will be taken by agents, that models will get cheaper and faster at incredible rates, and that modalities will integrate seamlessly. He tells builders to imagine everything being roughly ten times better in a year. He points to Notion, whose MCP drew a lot of agent traffic and strained the system, and says products must be built for scale and rethink the economics.
He changed his mind about timelines, since the capabilities of today's models arrived faster than he expected, and about reliance on voice and dictation. He also revised his view of younger generations, who embrace change first and absorb it quickly.
On safety, Sottiaux says pacing the frontier means investing ahead in alignment, security and guardrails, including compute for secondary monitoring that watches the primary agent for risky actions or prompt injection. He says OpenAI has not released the next capability step beyond Astra and had a model it chose not to release. The company serves 1.2 billion people, he says, and cannot gamble, and his remaining annoyance is the model picker, which he wants to make disappear.
More from this channel
1:10:27EN
Lenny's PodcastPodcast · 2 speakersHard Skills, Soft Skills and Mentors: Jules Walter's Playbook for Growing as a PMGuest Jules Walter · Host Lenny RachitskyYouTube product lead Jules Walter explains how he builds IQ and EQ skills, asks for honest feedback, and recruits mentors with the smallest possible ask.
1:30:46ENLenny's PodcastPodcastSkip the Funnel of Doom: Cursor's Playbook for Hiring the Top 50Host Lenny RachitskyAdam Ward, Cursor's head of talent, explains how to scope, map and relentlessly pursue the best people instead of hiring whoever survives a recruiting funnel.
1:42:04ENLenny's PodcastPodcast · 2 speakersPaid Growth, Step by Step: Shopify's Performance Marketing Lead on Tests, Signal and Team DesignGuest Timothy Davis · Host Lenny RachitskyTimothy Davis walks through when to start paid, how to run signs-of-life tests, which platforms to try, what to measure and who to hire first.
1:21:44ENLenny's PodcastPodcastFrom Rowing to Steering: Inside OpenAI's Push Toward Persistent AI CoworkersHost Lenny RachitskyTara Seshan, who leads product for Codex and ChatGPT Work at OpenAI, explains how PM work changes when agents do the rowing and people do the steering.
1:01:00ENLenny's PodcastPodcast · 2 speakersControl Over Attention: How Substack Builds Product for Writers and ReadersGuest Sachin Monga · Host Lenny RachitskySubstack's head of product, Sachin Monga, explains the principle behind the company's roadmap, the story of the Recommendations feature, and what he thinks the next era of writing on the internet looks like.
1:26:22ENLenny's PodcastPodcastStart Small, Earn Autonomy: What 50+ AI Product Deployments Taught Two BuildersHost Lenny RachitskyAishwarya Naresh Reganti and Kiriti Badam explain why AI products need a different build process, and how to grow agent autonomy step by step without losing customer trust.