TL;DR
The short version
OpenAI's Tara Seshan and Nan Yu tell Claire Vo to stop year-scale fantasy planning. Aim products at where models will be in two to three months, ship imperfect features to get empirical signal, and treat absorption — not raw capability — as the real constraint.
From the Lenny & Friends Summit panel: last-mile completion beats 99%-and-fail, PMs earn research trust with specific sessions and evals, and DMable user contact becomes table stakes for subtle agent failures.
Tara Seshan and Nan Yu (OpenAI), moderated by Claire Vo at Lenny & Friends Summit, 10 Sep 2026; published on Lenny's Podcast 29 Sep 2026.
Perfect a priori loses to empirical shipping
Seshan's Stripe polish habit met OpenAI urgency. The ChatGPT agent "toggle" is imperfect on purpose — a coherent step A so users can follow to step B. Quality still has a bar: additive value, internal uptake and delight, and alignment with model capability two to three months out.
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
Absorption, last mile, and layered platforms
Nan Yu names capability overhang: models outrun what people can absorb. Enterprises feel overwhelmed; shipping frontier agents anyway is doing what users need, not what they say. Computer use wins when it finishes the job — 99% then fail feels worse than a clean non-starter. Platform design is layers: hooks and plugins first, computer use as fallback.
Plan ~90 days. Stay DMable.
Seshan's planning horizon is two to three months — year-out predictions go in a personal dock of misses. Annual plans still fit slower markets like payments; OpenAI's market does not. What intensifies: onboarding under overhang, predictable agent behavior for privacy, faster empirical testing, and DMable PMs who can explain subtle agent failures to research.
