At OpenAI's 2026 DevDay, Lenny Rachitsky asked ChatGPT and Codex leader Tibo Sottiaux what comes after a better chatbot. His answer is an assistant that keeps a responsibility moving between conversations, uses the right tools, and interrupts only when a person needs to decide. The interview is most useful when read on two levels: what Dots can do now, and what Tibo expects the next interface to become.
OpenAI's Dots are real, gradually rolling-out persistent agents with a cloud computer, connected tools and a way to reach users across conversations. A future in which one agent absorbs every app, most internet actions come from agents, and users stop managing model pickers is Tibo's forecast. For builders, the immediate work is less glamorous: clear task interfaces, permission boundaries, dependable results and human review.
Watch the Interview
Source: Lenny's episode and transcript. The grouped chapter guide below is based on the supplied transcript. The official Dots launch article also includes OpenAI's launch and safety videos.
Dots Today: An Ongoing Responsibility, Not Just Another Chat
OpenAI's current Dots guide describes a persistent agent that can research, prepare files, work in its own cloud browser and bring results or decisions back to you. It can message or call through ChatGPT, and connect to Slack or Teams. Chat, ChatGPT Work, Codex and a Dot still have distinct jobs: conversation, delegated tasks, software work and an ongoing responsibility respectively. Tibo says OpenAI intends to make the boundaries feel simpler over time; that does not mean those products have already become one interface.
A Dot's cloud computer is separate from your laptop. Connecting a local computer or a plugin is a distinct permission step, and adding a messaging channel does not automatically grant access to your inbox or files. As of 26 October 2026, OpenAI says Dots are rolling out gradually: eligible personal Pro users outside the EEA, UK and Switzerland, Business Premium users worldwide, and Enterprise workspaces whose administrator enables them. Availability may differ by plan, region and workspace. That regional restriction matters for readers in Portugal. Check the live access guide before buying a plan for this feature.
Tibo describes starting with one primary Dot and later adding specialist ones. OpenAI's launch post similarly presents teams of Dots as a future direction, while enterprise specialist Dots are a preview. Do not design a current workflow around unlimited personal Dots yet.
Why Tibo Wants the App to Fade Into the Background
In the interview, Tibo says his own agent use expands into several parallel workers, then contracts when a stronger model can keep more of a problem in context. That is why he is skeptical that hand-tuned agent loops and graphs should be the permanent user experience. His preferred end state is an assistant that learns a person's goals and standards, chooses a way to proceed, and stays available through voice or the right screen without making the user configure every step.
This is a product thesis, not evidence that orchestration, workflows or model selection no longer matter. Under the hood, reliable systems still need task definitions, evaluation, recovery and controls. The sensible translation for a team today is to hide incidental configuration from end users while preserving observable steps and clear approval points for consequential actions.
The Quiet Bet: Plugins, Distribution and Shared Economics
Tibo calls the ecosystem the underappreciated DevDay story. OpenAI's official recap confirms plugin extensions, improved submission and discovery, and Sign in with ChatGPT usage at participating partners. The plugin guide explains that plugins can bundle skills and connected tools, but each external account and permission still needs review.
In the interview, Tibo says OpenAI intends to share economics with popular plugins and use quality, task success and retention in recommendation decisions. Those are his statements about platform direction; the public DevDay recap documents discovery improvements but does not provide a universal payout formula or guarantee that a particular plugin will be promoted. Builders should validate a real repeated task and a sustainable integration before treating the marketplace as a distribution shortcut.
| For a plugin builder | What to test first |
|---|---|
| Useful task | Can a user complete one concrete job with fewer steps and a checkable result? |
| Permission design | Does it ask only for the data and actions the job requires? |
| Reliability | What happens when the service is slow, unavailable or returns partial data? |
| Return value | Do users come back for that job without needing a promotional prompt? |
What Changes If Agents Become a Major Customer?
Tibo predicts that a majority of internet actions will eventually be taken by agents. That is his forecast, not a present-day traffic statistic. His concrete example is an integration that makes a product usable by many agents and can suddenly increase load. Product teams would then need to consider rate limits, authorization, cost per action and support for agent failures, not just page design.
For a website or SaaS product, the useful near-term test is modest: make pricing, availability, service details and policies legible; expose a reliable path to an authorized action; and record whether real customer agents can complete it. A full agent API is not automatically required. Our browser-versus-API decision guide sets out the trade-off. Tibo also argues for better human experiences, particularly multimodal collaboration, so optimizing only for machine traffic would miss half of his point.
What Becomes More Valuable for People?
When Lenny asks which skills are moving up, Tibo names taste, understanding the user, collaboration and knowing what good looks like. Fast typing becomes less distinctive. He praises a colleague's kindness, ability to learn quickly and willingness to work on the most important problem. This is an interviewee's hiring perspective, not a labor-market forecast that coding or engineering fundamentals no longer matter.
The conversation also acknowledges the less celebratory side of agents: context switching, isolation and pressure to do more simply because more can be produced. Tibo's response is that useful AI should reduce noise and protect attention. The test for any agent at work should therefore include time and cognitive load saved, not just tasks initiated or lines generated.
The Outage Anecdote, and the Boundary It Illustrates
Tibo tells Lenny that his Dot warned him of a production outage about five minutes before a live demo, connecting the outage to his calendar and the event. It offered to fix the problem; he declined and contacted the engineering team. This is a single reported anecdote, not proof that Dots reliably detect outages or should receive production-write access. The valuable behavior was identifying a relevant signal and escalating it to the right human.
OpenAI's launch article says background proactive research uses read-only app access, while follow-up actions are subject to permissions and review. Its controls guide details activity review, approvals and custom rules. Tibo also discusses monitoring and alignment in the context of more autonomous systems. OpenAI's own Hugging Face incident report describes why stronger containment and escalation became necessary. None of these safeguards remove the need to review sensitive outputs and tightly scope access.
A Good First Test, If Your Account Has Access
- Choose one ongoing responsibility. A project deadline or weekly status is better than “run my business.”
- Give it only the necessary sources. Prefer a low-risk project and read-only access at first. Confirm the connection can see the intended workspace, not everything by accident.
- Define the interruption rule. Ask for routine progress in one channel and an alert only when a deadline, blocker or decision changes.
- Keep actions reviewable. Have it draft a message or document for approval before sending or publishing.
- Measure a week of outcomes. Count useful alerts, false alarms, work corrected and actual minutes saved. Remove connections that did not earn their access.
This is an editorial pilot derived from the interview and OpenAI's guidance, not a promise that every account has Dots or that a Dot can complete any particular integration today.
Video Chapters by Theme
Jump to the conversation that matters most. These links preserve the interview's order within each topic group; times are from the supplied transcript.
Interview and product vision
| Time | Discussion |
|---|---|
| 00:00 | Introduction and the agent-era thesis |
| 01:12 | How Tibo uses AI at work |
| 03:10 | A persistent agent instead of manual loops |
| 06:15 | Why work still feels clunky |
| 07:32 | Bringing Codex, ChatGPT and Dots together |
| 09:00 | Building with the community |
Dots and the developer ecosystem
| Time | Discussion |
|---|---|
| 10:13 | One primary Dot now; more are planned |
| 11:34 | Open ecosystem, partners and plugin economics |
| 13:39 | How plugins might be recommended |
| 14:39 | The research behind persistent agents |
| 20:09 | An outage alert before a live demo |
| 21:42 | Guardrails and connected devices |
People, strategy and safety
| Time | Discussion |
|---|---|
| 16:20 | Where human judgment remains valuable |
| 17:25 | Agent fatigue and context switching |
| 23:25 | Skills gaining value |
| 25:56 | Advice for new graduates |
| 27:18 | How OpenAI teams build and ship |
| 30:45 | A forecast: agents dominate internet activity |
| 33:55 | Safety, monitoring and alignment |
| 36:15 | Moving beyond model pickers |
Sources and Scope
The primary account is Lenny's interview and its episode transcript. Product and rollout details were checked against Introducing Dots, the Dots guide, DevDay 2026 recap, plugin documentation and OpenAI's incident report on 26 October 2026. Tibo's predictions, outage story and plugin-revenue comments are attributed to the interview; we did not independently test Dots, its alert quality or plugin payouts. Follow Tibo on X and Lenny's Newsletter for their own updates.