OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux

Tibo Sottiaux, introduced as OpenAI's head of ChatGPT, joins Lenny's Podcast to discuss persistent assistants, agent infrastructure, and the product experience he wants to build. His central direction is less configuration for users: an assistant that remembers goals, works in the background, and can be reached through different devices and channels.

Evidence basis: This summary uses the saved English auto-generated captions, without an original-audio check. Product capabilities, usage figures, and operational examples are Sottiaux's first-party accounts; roadmap statements and predictions are not independently verified results. Ambiguous caption renderings of model names and numerical claims are omitted.

Persistent assistants and changing orchestration

• More capable models can reduce the need for elaborate agent teams. Sottiaux says he expands his agent team when pushing what current models can do, then shrinks it after a breakthrough lets one stronger agent handle more work. Faster responses also help him regain a sense of flow. This is his workflow experience, not a general benchmark showing that single-agent systems always outperform multiple agents. 01:44–03:42(opens in a new tab)
• He wants users to specify goals rather than maintain execution loops. His proposed alternative is a persistent assistant that understands preferences and learns from feedback. He acknowledges that the initial product is imperfect. The interview does not explain the learning mechanism or show that explicit orchestration is unnecessary for every task. 02:38–04:44(opens in a new tab)
• Reducing configuration is a product goal, not a completed transition. He describes merging chat and work experiences and eventually bringing persistent-assistant capabilities into ChatGPT. Voice and shared collaborative surfaces would let people work with other humans while an agent builds in the background. He explicitly says that experience is not fully realized yet, and closes by criticizing model pickers, reasoning-effort controls, and multi-agent configuration fatigue. 06:51–08:40(opens in a new tab), 17:13–18:55(opens in a new tab), 36:15–37:08(opens in a new tab)

Distribution and infrastructure for agent use

• Plugin discovery is described as a utility-and-retention feedback loop. Sottiaux says OpenAI recommends plugins using signals such as quality, successful use, and retention, and stops recommending poor ones. He also describes shared economics for plugins and products using ChatGPT sign-in. These are his account of the platform direction, not documented payout terms or a guarantee of distribution. 11:20–14:15(opens in a new tab)
• An agent's coordinating software need not live on the device it controls. He says the Dots harness runs separately from connected machines and can reach multiple devices, rather than being tied to one computer. He describes internal specialist agents with extra guardrails, monitoring, and dedicated hardware, including Mac minis. The interview does not specify the permission model or deployment architecture in enough detail to reproduce it. 21:53–22:49(opens in a new tab)
• Agent-facing interfaces can change traffic volume and costs. Sottiaux cites Notion's Model Context Protocol (MCP) integration as an example of exposing a product to agents and then confronting increased traffic and operational strain. His forecast is that agents will perform most internet actions; his practical concern is whether services can support the resulting scale and economics. He also argues for investing in better multimodal experiences for human users, rather than treating agent access as the only interface problem. 30:47–33:41(opens in a new tab)

Awareness is not authority to act

• A useful alert did not justify an autonomous production repair. Sottiaux reports that his assistant connected a production outage with an imminent DevDay demonstration and warned him. When it offered to fix the outage, he declined and contacted the engineering team instead. The anecdote illustrates useful contextual awareness alongside a human decision to withhold an action; it does not establish reliable incident diagnosis or remediation. 19:57–21:53(opens in a new tab)
• Monitoring is described as a separate investment from task execution. He distinguishes compute spent on the agent doing work from compute spent watching for risky actions or possible prompt injection and intervening. He also discusses withholding a more capable model release while investing in safety and security. These are descriptions of OpenAI's approach, without detection rates, failure analysis, or evidence that monitoring eliminates the risks. 33:41–36:15(opens in a new tab)

Human judgment and attention

Sottiaux emphasizes taste, understanding users, collaboration, and rapid learning over typing speed. His example of a successful early-career hire stresses kindness, solving important problems, and learning broadly. On pressure to run ever more agents, he argues that AI should reduce competing demands on attention and leave people better rested—not merely increase prompting throughput. He presents that as an industry problem still to solve. 18:55–20:51(opens in a new tab), 23:44–24:31(opens in a new tab), 26:06–27:59(opens in a new tab)