Tibo Sottiaux, Head of ChatGPT at OpenAI, discusses the rapid evolution of AI agents, the launch of “Dots” (always-on autonomous agents), and how OpenAI is building toward a future where the majority of internet actions are performed by agents rather than humans — a shift he believes most builders are still underestimating.
Tibo’s daily workflow with AI agents
Tibo no longer writes code by hand; nearly all analysis, trend detection, and feature exploration code is generated by Codex.
He occasionally does “therapeutic” hand-coding on weekends, but day-to-day work is delegated to agents.
The number of parallel agents he runs fluctuates: as models improve, a single more capable agent replaces teams of smaller agents, creating an expansion-and-contraction cycle.
With the recent “ultrafast” model breakthrough, he’s back to running fewer, faster agents and staying in flow.
Vision for always-on agents (Dots)
Dots are designed to be permanent, active intelligence that works 24/7, understands goals and preferences, learns from feedback, and is accessible from any device or screen — no laptop required.
The goal is to break free from being “tied to the technology” and have the technology work for you, appearing in meetings, email, text, or voice as needed.
Tibo sees this as the primary interface for AI, replacing the need to choose between Codex, ChatGPT, or other tools.
Merging Codex, ChatGPT, and Dots
OpenAI is merging the “work” toggle (Codex capabilities) directly into ChatGPT to reduce complexity; users loved Codex’s power but preferred ChatGPT’s speed and UX.
All Docs capabilities will eventually ship into ChatGPT for the full user base (1.2B+ users).
Dots have no model picker and minimal configuration — you only choose which channels (Slack, email, etc.) the agent connects to.
Sci-fi inspiration
Tibo draws more from 30–40-year-old sci-fi: Neuromancer, original Star Trek — the vision of talking to a computer that just does things, like commanding a spaceship.
Her was watched once but isn’t the primary reference point.
Building with the community
Tibo emphasizes that OpenAI discovers what the technology is by launching and watching how people use it; community feedback is essential to building something truly useful.
He spends ~30 minutes/day on Twitter, engaging directly with users; sees it as a privilege to stay close to the community without approval layers.
Dots structure: primary dot and team dots
Launch starts with one “primary dot” per user — the main agent that learns preferences deeply and connects to apps.
Soon users will create multiple specialized dots (a “virtual team”) for distinct roles: e.g., a Twitter-monitoring dot, a research dot, a coding dot.
Each dot can be given specific responsibilities and guardrails.
Open ecosystem and plugin revenue sharing
The “sleeper hit” of the launches is the open ecosystem: 16 launch partners for “Sign in with ChatGPT,” plus plugin extensions and discovery.
OpenAI will share revenue with popular plugins that drive usage — when subscribers use a plugin, the plugin creator gets a cut.
This mirrors the early handshake deals with Pi and OpenCode creators, now formalized at scale.
Plugin discovery and recommendation
Discovery is based on retention and quality signals, not SEO or marketing: plugins that users keep using get recommended in conversations.
If a plugin stops being useful, recommendations stop — a meritocratic loop.
Research behind Dots
Work on long-horizon tasks, persistent memory, and 24/7 operation spans >2 years.
Memory systems (launched in ChatGPT) already surprise users with recall of personal details from months prior.
Dots combine the Quora harness, memory, and the ability to run continuously with good productivity.
Safety was a major factor: Dots launch on Astra (OpenAI’s safest, most aligned model) with extensive guardrails.
Where humans remain valuable
OpenAI designs technology to put humans at the center — as extensions of human will, taste, and creativity.
Humans will remain the ones who decide what to build, what “good” looks like, and who want to see what other humans create.
“We may not have coders anymore, but we have more builders than ever.”
Agent fatigue, loneliness, and context switching
Engineers report increased loneliness and context switching from managing many solo agent conversations.
Solutions in progress: reduce configuration fatigue, make agent collaboration feel natural (shared whiteboards, voice, physical-space presence), and move beyond screen-bound prompting.
The ideal: agents observe conversations, jot ideas, build in background, project on screens — all while humans talk to each other.
Pressure to do more
AI creates pressure to ship more, run more agents, attend more meetings.
The promise of AI should be reducing noise, letting people focus attention where they choose, and enabling better rest — not just “one more prompt per second.”
Tibo finds he thinks more creatively after a full disconnect; the goal is to bring that state into daily work.
Agent caught a production outage before a live demo
A personal “energy audit” dot monitoring Tibo’s calendar and systems detected a ChatGPT production outage 5 minutes before his DevDay live demo.
The dot understood the context: DevDay event, live demo dependency, production system — and proactively alerted him, even offering to fix it.
The dot had guarded access to some production systems (not full codebase) via specialist-dot guardrails and dedicated hardware (Mac minis).
Agent guardrails and multi-device control
Specialist dots run with additional guardrails, monitoring, and on isolated hardware.
The agent harness doesn’t run on the user’s machine; it lives in a VM/cloud and can connect to many devices (laptop, phone, servers) simultaneously — “like an octopus.”
This architecture enables secure, scalable delegation.
Skills trending up and down in the AI era
Trending down: typing speed, pure implementation mechanics.
Trending up: taste, user empathy, connecting with the audience you’re building for, knowing what “good” looks like.
Ex-founders thrive: OpenAI has >120 YC founders; the “passion to build something that matters” matters more than ever.
PM skills (deciding what to build, iterating, judging quality) are increasingly valuable.
Roles are blurring: designers who code, engineers who design — “your time to shine.”
Tibo’s return to engineering flow
Missed hands-on engineering for a while, but ultrafast models + voice control restored a new kind of creative flow state.
Different from the old flow, but equally immersive; excited to democratize this for billions.
Advice for early-career professionals
The “next Tibo” is already at OpenAI: Ahmed Ibrahim, hired as a new grad, now owns the entire compute fleet and applied team.
Ahmed’s edge: extreme kindness, collaboration, solving the most important problems (not self-promotion), and being a “sponge” — learning at unprecedented speed using AI tools.
Junior people have an advantage: no legacy workflows to unlearn; they go all-in on AI-native work.
How teams build and ship inside OpenAI
Bottom-up, founder-energy culture: small Slack channels form, 4 people hack on a weekend, then it opens up company-wide, excitement compounds, more people pitch in.
Example: Decisions API shipped fast after realizing Luna model + constraint sampling could create a faster, structured-output API.
Leadership holds a high quality bar; many things were held back from DevDay for more baking — launches spaced out over following weeks.
Chaos channeled into productive shipping through informal role assumption.
Autonomy, trust, and learning from mistakes
High autonomy paired with ownership: “you break it, you fix it, you learn.”
Tibo took down production on day 3 at OpenAI; still employed, learned the lesson.
Autonomy lets him stay close to community without approval chains.
Mistakes happen: Tibo admits he sometimes pushed teams toward over-complexity early in Codex; caused a couple outages.
Building for an agent-dominated internet
Three under-priced shifts: (1) majority of internet actions will be by agents, (2) models will get cheaper/faster at incredible rates, (3) all modalities will integrate seamlessly.
Most builders are “not quite getting it” — if they imagined 10x better in a year, they’d build differently.
Second-order effects: products must build for agent-scale traffic (Notion saw massive MCP-driven traffic); economics of agent usage create tension (interface vs. API); delightful human-first multimodal experiences are under-invested.
What Tibo changed his mind about (past year)
Expected Astra-level capabilities in 1–2 years; arrived much faster.
Didn’t anticipate how much he’d rely on voice/dictation — it feels dramatically better than expected.
Underestimated the talent and speed of younger generations entering the workforce; they embrace AI-native workflows first and drive hiring strategy shifts.
AI safety and alignment approach
“Pacing the frontier” = investing far ahead in alignment, safety, security, guardrails.
Majority of API-stack compute goes to secondary monitoring: watchdog models oversee primary agents for risk, prompt injection, unsafe actions.
Next capability step beyond Astra not yet released; current releases are near-Astra intelligence but more efficient.
Proud of not releasing “6.1 Astra” when it wasn’t ready — safety over speed.
Incentives align: 1.2B users (many non-technical) means OpenAI cannot afford to screw up safety.
Moving beyond model pickers toward simplicity
Current annoyance: model picker, reasoning effort toggles, multi-agent vs. ultra choices — “you need a PhD in model pickers.”
Goal: make the app almost disappear into a “super essence of simplicity” where the right model/tool is selected automatically.
Tibo himself gets fatigued by the current complexity; wants it gone as fast as possible.
