Shipping the Imperfect: OpenAI's Seshan and Yu on Building AI Products for Tomorrow
Key points
Key takeaways from the Lenny's Podcast conversation with OpenAI's Tara Seshan and Nan Yu, hosted by Claire Vo, recorded live at the Lenny and Friends Summit in San Francisco on 10 September 2026:
Ship imperfect, ship now. Seshan, who spent some seven years at Stripe, says urgency and empirical evidence from real users now beats a priori polish. The "toggle" was an imperfect solution, but it got the agentic harness into the hands of more than a billion ChatGPT users without breaking developers' workflows.
Build for where the models will be in two to three months. Her quality bar combines additive value, something users care about, surprising delight, and alignment with near-term model capabilities. Neither anchored in the present nor futuristic-and-unusable.
The binding constraint is users' ability to absorb change. Yu argues models have "capability overhang" and people do not take advantage of what is already possible —so it does not matter what you ship if users cannot absorb it — yet not shipping the frontier gets you leapfrogged, even when that means breaking enterprise processes.
Do what your users need, not what they say they need. OpenAI had to push agents into enterprises faster than their own change cycles so they would not defect to alternatives — a truism of product that applies "almost more than it ever has before".
Agents map onto human nature. On single-identity versus many agents, Seshan says most people cannot manage 40 threads. The most successful designs group activity into bundles — such as a chief-of-staff agent that manages all the other ones for you.
Tactical edge cases decide agent design. Yu argues the philosophical split matters less than the use case: data access, permissions, segmented memory and credentials are what actually determine whether one unified agent or several micro-agents is right.
Two crafts plus relentlessness. Seshan describes agent design as user empathy ("inception" — understanding mental models) plus systems thinking ("severance" — edge cases and platform components), alongside a relentlessness to keep trying and iterating at a much faster clock speed.
The last mile decides whether a capability feels great. A task that gets 99% done and then fails is worse than a non-starter, she argues. Computer use feels great because it always gets the job done all the way — making the user feel like a guest in your home, with needs anticipated.
Read more: Lenny's Podcast on YouTube
