OpenAI is eliminating model selectors: the future is one resident agent
OpenAI is merging Codex, ChatGPT, and Dots into a single resident agent without a model selector. The goal: users never choose models, never juggle multiple agents, just state what they want to do.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Dots is a resident agent without a model selector
Tibo says that regardless of the name, the core is breaking free from being ‘pasted onto a specific client’: a resident agent that's always online, knows all your goals and preferences, and can be invoked from any screen—taking notes in meetings, following up on emails and messages. You don't keep watching your laptop or phone; it appears when needed and disappears when not, rather than dragging users around like a brick.
— Tibo SottiauxCodex and ChatGPT will merge, capabilities extended to 1.2 billion users
Tibo says Codex and ChatGPT's chat and work modes will be unified, eliminating users' burden of choosing between ‘fast but limited’ and ‘powerful but slow’. All Dots capabilities will integrate directly into ChatGPT to ‘lift the floor’—raising the baseline experience for 1.2 billion users—while Dots itself deliberately has no model selector; the only configuration is which channels to connect.
— Tibo SottiauxThe plugin ecosystem is the real hidden blockbuster in today's launch
When asked which announcement was underestimated, Tibo pointed directly at ecosystem opening: last year's ‘Sign in with ChatGPT’ began as mere ‘virtual handshakes’ with founders like Perplexity and OpenCode, then evolved into formal multi-partner integration. Now plugins can reach 120 million users' distribution channels, and popular, high-usage plugins will receive revenue sharing—something not even mentioned in the keynote.
— Tibo SottiauxAn agent caught a production incident five minutes before going live
Before the Dev Day broadcast, Tibo's dot recognized it was launch day, ChatGPT production had a system issue, and the event had a live demo—these things were connected—and warned him five minutes early that ‘production is down’, even offering to help. Tibo's response was ‘thanks, but I don't think you're at that level yet’, then he contacted the engineering team. He sees this kind of cross-scenario correlation as genuinely impressive.
— Tibo SottiauxDots doesn't run on your device—it connects multiple machines like an octopus
The fundamental difference is that the execution harness doesn't run on your own device; instead it can simultaneously connect any number of machines—your laptop, or possibly dedicated hardware like a Mac mini. Specialized ‘expert dots’ handling sensitive tasks run with extra guardrails and monitoring on isolated hardware. In theory, a single dot could eventually control ten devices at once, like an octopus.
— Tibo SottiauxFast typing is becoming worthless; founder mindset and taste are gaining value
Tibo says fast typing as a skill is becoming obsolete, while taste and user-first thinking are gaining value. Many former founders and serial entrepreneurs now stand out; OpenAI currently has over 120 former founders on staff, which he calls ‘a mega startup’. Role boundaries are blurring too—previously being design-only or engineering-only felt limiting; now is the time to ‘shine’.
— Tibo SottiauxMost internet traffic will soon come from agents—products need to handle the load now
Tibo believes the notion that ‘most internet behavior will be performed by agents’ hasn't been properly priced in yet. He cites Notion: after building MCP, suddenly large numbers of agents poured in making calls and doing actual work, traffic exploded, and the system faced real pressure and economic consequences. His conclusion: whether to build a human-usable interface can wait, but scaling for agents is inevitable and most products haven't truly thought it through.
— Tibo SottiauxSafety investment: monitoring agents costs as much compute as the main agent
On AI risks, Tibo says ‘staying at the frontier’ for him means investing early and investing more than he thinks ‘should’ be necessary: alignment, safety, and guardrails get continuously increased investment. Compute dedicated to ‘second-order monitoring’ of the main agent's behavior is growing substantially—checking whether it's taking high-risk actions or showing signs of prompt injection. He also mentioned the team built Astra 6.1 but chose not to release it, a decision he's proud of.
— Tibo SottiauxIn their own words · checked verbatim
it pinged me five minutes before the live demo, which was actually quite stressful because it was like, hey, production is down.
Tibo Sottiaux20:09
Over time, you might connect it to like 10 different devices and it can control it all a little bit like an octopus.
Tibo Sottiaux21:36
Yeah, I would say like a skill that is trending down is typing fast. That is not that useful anymore.
Tibo Sottiaux23:39
Very, very early days for me at OpenAI, I took production down like day three.
Tibo Sottiaux29:43
I think the majority of actions on the internet will be taken by agents.
Tibo Sottiaux30:45
We've worked very closely with Notion, for example, and when they build their MCP, suddenly it's like, oh, well, it's available to all these agents that can actually do the work and use that MCP.
Tibo Sottiaux31:43
I myself even get fatigued with the model picker and the reasoning efforts and, you know, whether to use multi-agent or ultra or, you know, what it even does. And you kind of need a PhD in model pickers.
Tibo Sottiaux36:52
Figures
| ChatGPT user scale | 1.2 billion | 7:19 |
| Former founders on OpenAI staff | over 120 | 23:39 |
| Time Tibo spends on Twitter daily | around 30 minutes | 10:20 |
| Number of devices a single dot could eventually connect | around 10 | 21:36 |
Glossary
- Dots
- OpenAI's new resident agent product, no model selector, configuration limited to which channels to connect
- MCP
- Model Context Protocol—standard connectivity protocol letting third-party apps and plugins integrate with ChatGPT and various agents
- harness
- Agent execution framework—the underlying scaffolding code that runs agents, calls tools, and decouples from specific devices
- prompt injection
- Attack where adversaries craft inputs to trick the model into executing unintended, unprivileged actions
- Astra
- OpenAI's internal code name for the safest, best-aligned model
How to listen
Product managers and founders wondering whether to include model selectors in products, how to structure plugin ecosystem revenue-sharing, and how much to invest in agent safety.
The section from 8:20-9:02 about sci-fi inspiration and community culture has low information density and can be skipped.