Stop planning for 2027: how OpenAI builds product 90 days at a time | Tara Sesha and Nan Yu (OpenAI)

Episode Overview

• Moderator Claire Vo speaks with Tara Seshan and Nan Yu about how product leaders build, ship, and iterate on AI products and agentic experiences when underlying models are evolving rapidly.
• The conversation explores the transition from deeply polished traditional software launches to urgent, iterative AI releases, highlighting how user comprehension has replaced model capability as the primary bottleneck.
• The speakers break down practical frameworks for designing single versus multi-agent architectures, overcoming the "last mile" problem in software integrations, and collaborating directly with AI research teams.
• Product managers, builders, and enterprise leaders will learn how to adapt their planning horizons, onboarding flows, and feedback loops to match the pace of frontier AI.

Key Concepts

• Urgency Over Polish in AI Releases: At the Stripe-sponsored Lenny & Friends Summit, OpenAI ChatGPT Work Product Lead Tara Seshan contrasts her 7 some odd years at Stripe, where products are deeply polished before launch, with the urgency required at OpenAI. She explains that shipping an imperfect UI toggle between standard chat and agentic work modes was necessary to get an agentic harness to "over a billion users that use ChatGPT" without disrupting developer workflows. OpenAI Codex Product Lead Nan Yu adds that the obvious next step is "no toggle," noting users accept product churn if guided with transparency and a coherent story.
• Capability Overhang and Absorption Limits: Nan Yu argues that the primary constraint in AI product development is no longer what the models can do, a phenomenon known as capability overhang, but rather users' ability to understand and absorb new features. Consequently, onboarding has become more critical than ever, and he predicts the rise of "self-driving" products next year that use their own intelligence to solve the "empty input box problem" and gently guide users.
• Pushing Enterprise Adoption: Conventional wisdom dictates that enterprise customers cannot handle rapid change. However, Tara Seshan claims that withholding frontier updates causes companies to get leapfrogged. Because most of OpenAI's enterprise customers were only using basic chat, OpenAI intentionally shipped a disruptive evolution toward agents that broke standard enterprise update processes so customers could unlock new value rather than look at alternative products.
• Human-Centric Agent Architecture: Moderator Claire Vo, who builds businesses relying on OpenAI tools and keeps a single Codex agent docked on her computer all day, contrasts her setup with having "40 Grok bots" across a business. Nan Yu explains that agent design must map onto human evolution; managing 40 agents is like managing 40 direct employees, which almost no one (except perhaps Jensen Huang) wants to do. He notes that users who claim to manage 7 agents usually "cheat" by creating 1 "Chief of Staff" agent to manage the other 6. Tara Seshan adds that choosing between single and multi-agent identities comes down to tactical product details: data access, user versus service account permissions, and whether an agent in a private Slack channel has "severed" or segmented memory.
• Solving the "Last Mile" with Computer Use: Tara Seshan describes ChatGPT as a layered platform where first-party or third-party plugins and hooks form the first layer, and computer use acts as a fallback layer. Nan Yu explains why this fallback matters: when an AI product gets 99% of a job done and "barfs" on the last mile, it feels worse to a user than a total non-starter. Even though computer use is slow and costs many tokens, it guarantees 100% completion while waiting for the broader ecosystem of MCPs and hooks to catch up.
• Shortened Planning Horizons in Frontier AI: Unlike established markets such as payments at Stripe where bull, bear, and base cases can be modeled many months out, Tara Seshan notes that OpenAI's business trajectory cannot be predicted the same way (though she hedges that perhaps OpenAI CFO Sarah Friar is the expert who can do so). Seshan argues that multi-year AI predictions are "almost always wrong" and notes she keeps a personal document of her own incorrect predictions as a reminder.
• Future Form Factors: Looking ahead to 2027, Tara Seshan is "so bullish on voice" (referencing GPT Live), noting it reduced the tech support she does for her family and transformed how she onboarded OpenAI's internal People team. Meanwhile, Claire Vo predicts new hardware form factors, such as a "baby bot" or "Codex in a box," that replace carrying around an open laptop.

Quotes

• At 1:26 - "And also, there's a lot of like theorizing that one can do before you ship something, but nothing compares to like the actual empirical evidence of seeing users try it, use it, and then iterating from there. So I've really had to shift my mindset from how do you get it perfect a priori versus how do you ship something and iterate on it as quickly and effectively as possible." - Tara Seshan explaining why empirical user feedback in frontier AI outweighs pre-launch perfection.
• At 5:15 - "The biggest constraint is honestly like people's understanding and ability to sort of absorb what you're what you're giving them, right? We talk a lot about like capability overhang, about like the models can do so much and just like, you know, people aren't necessarily taking advantage of all those abilities. I think that's where that's where the gap is, right?" - Nan Yu identifying user comprehension as the primary bottleneck in AI adoption.
• At 15:50 - "There is a huge difference between getting like all of a job done versus getting everything except for the last mile. And the last mile honestly feels sometimes worse than just like, oh, it's a non-starter. If it's a non-starter, you can just like, look, I'm just going to do it myself or I'll try another route or whatever it is. But if it gets like 99% there and then it just like barfs, then you're like, well, this is actually a worse sort of like unfulfilled promise." - Nan Yu illustrating why reliable fallback mechanisms like computer use are essential for user trust.
• At 18:21 - "In as much as like I can add something to working with research, it is for me to come with very, very specific use cases, a really clear idea of what like users want to do and their goals, to have done like sample reading, so understand like literally what was the like session and what did it look like and why did the model not do what it needed to do. If I can write evals, that is like I'm there, like learning how to write good evals and writing evals as much as possible is like the most important thing." - Tara Seshan detailing how product managers practically contribute to AI research teams.

Takeaways

• Anchor product roadmaps to a 60-to-90-day horizon: Evaluate new features using three checks: ensure they unlock genuinely additive value, pass a qualitative internal dogfooding bar for retention and delight (which Seshan notes is not a specific number or exact metric), and aim at where models will be in 2 to 3 months so you are "sufficiently AGI-pilled" and getting out of the model's way without building for today's limitations.
• Contribute to AI research via sample reading and evals: While hedging that she is not an expert on collaborating with researchers and offering a "novice's view," Tara Seshan advises product managers to bring specific user goals to research teams, read raw session logs ("sample reading") to diagnose exact model failures, write evaluations (evals), and prove out outcomes with prompts and skills first so post-training teams can bake those capabilities directly into the model.
• Be a "DM-able PM" and clarify product norms: Maintain direct, accessible relationships with end users, even on massive consumer products, because subtle natural language bugs (like an agent "giving attitude") require 2 or 3 follow-up questions to diagnose; additionally, design clear data privacy and product norms so users can accurately predict what semi-autonomous agents will do at any given point.