Plan in 90-Day Windows: How OpenAI's Product Leaders Ship Fast Without Losing the Plot

Host Claire Vo opens with the "toggle", the fact that every AI product has a left and right setting, and asks how OpenAI decides to ship something imperfect. Seshan says that after about seven years at Stripe, where everything was polished and every enum name considered, she has had to change her mindset. Urgency now matters, and nothing compares to watching real users try a feature and iterating from there.

She says the toggle in question is certainly not the ideal solution, but it was important to get an agentic harness into the hands of the more than a billion people who use ChatGPT without disrupting developers' workflows. Yu adds that the obvious next step is no toggle at all. If users can follow a coherent story from step A to step B, with some transparency into what is going on behind the scenes, they accept the evolution and bridges are not burned.

Asked what keeps the quality bar high, Seshan lists a few tests: is it additive and genuinely useful, and does it get uptake and retention when used internally, with something delightful or a new use of the models? Most important in this era is whether it targets where the models will be in two to three months, and whether it gets out of the model's way.

Yu says the biggest constraint is people's understanding and ability to absorb what they are given. He points to capability overhang: models can do much more than people are taking advantage of, so the gap lies in adoption. Vo notes that B2B customers often say they cannot absorb change at this pace.

The guests concede that to most enterprises the current pace feels beyond breakneck, but say that not shipping the frontier means getting leapfrogged. Most enterprise customers were merely asking questions in chat while the agent revolution was happening, and OpenAI needed to bring agents to them quickly even though it broke their processes and expectations about how updates arrive. They call this the old product truism turned up: do what users need, not what they say they need.

Vo asks whether the future is one generic agent identity or many micro agents, as she runs a Codex she talks to all day and around 40 specialized bots. Yu says the most successful designs map onto human nature: 40 agents is a lot to hold in your head, and people who say they manage many usually have a chief-of-staff agent managing the rest. Seshan adds that practical questions decide it, including data access and permissions, whether an agent uses your credentials or a service account, and how memory behaves in a private Slack channel.

On skills, Vo describes a human side (mental models, empathy) and a systems side. Yu calls these the old product principles in disguise, and Seshan adds a third: a relentlessness to try, iterate and learn from feedback loops, now at a much faster clock speed. On the ecosystem, one guest describes a platform exposing layers of hooks for first-party and third-party developers, with computer use as the fallback when pre-built tools are missing.

One guest says there is a huge difference between doing the whole job and getting everything except the last mile: an agent that gets to 99 percent and then fails is worse than a non-starter, an unfulfilled promise. Computer use may be slow and costly in tokens, but it will do the job all the way while the ecosystem catches up with the right hooks and MCPs. The goal, quoting Charles Eames, is that the user should feel like a guest in a home where their needs were anticipated.

Seshan says working with research was entirely new and is different from working with engineering. Her contribution is to bring very specific use cases, clear user goals and sample sessions showing why the model failed. Writing evals is the most important thing, and the ideal is to show that a given prompt and skills achieve the outcome so research can try to post-train that capability into the model.

On time horizon, Seshan says the ideal is two to three months, since predictions years out are nearly always wrong, although building only for today means being left behind. Vo polls the audience on who is doing 2027 annual planning. Seshan notes that it depends on the business: Stripe's payments dynamics can be modeled with bull, bear and base cases, whereas OpenAI's cannot.

On what still matters, Yu says onboarding matters far more now because of capability overhang, and Seshan says users must be able to predict what an agent will do and understand how their data is used. The usual PM craft still applies, but testing empirically and moving faster has become more important. Both argue for direct, DM-able relationships with users: agent problems like a model that came across as having an attitude are subtle and need follow-up questions.

Asked what will be big in 2027, Seshan answers voice. She says it has changed how she works, feels more natural, and has cut the tech support she does for her family, and that it made onboarding OpenAI's own people team onto a new product much easier. She ties this to the idea that human nature, evolved for speaking with others, makes voice intuitive now that the models are finally getting good at it.

Yu predicts something that looks like self-driving. He says the empty input box problem, where you have the product and now must learn how to use it, will be solved by products that are intelligent enough to use themselves and give a gentle on-ramp. Vo adds her own guess of hardware that replaces carrying a laptop, and they agree to see who was right.