Tara Seshan, who leads Codex and ChatGPT Work at OpenAI, says the third era of AI products is a persistent coworker agent that gets things done.
Seshan argues product teams should build for the models expected in two to three months, not for today's models or next year's.
ChatGPT's separate chat and Work modes are a temporary split, with the goal being one experience without a toggle that brings agents to about a billion chat users.
OpenAI's Tara Seshan on building products for AI coworkers
AI at work is shifting from chat toward persistent agent coworkers that do the execution while people set direction. Because models improve so fast, product teams have to build for the models expected in two to three months and learn through quick, empirical testing with users. Seshan's view is that the winning move for people is to aim higher rather than merely automate what they already do.
Takeaways
The third era of AI is persistent coworkers. Seshan says AI moved from chat to agents, and the next era may be working alongside a persistent coworker that gets things done with people. ChatGPT's chat and Work modes are a temporary split, and the goal is one experience with no toggle.
Build for models two to three months out. The field is emergent and changes fast, so teams cannot build for today's models or next year's. Seshan asks builders whether they are building for where models will be in two to three months, echoing that current models are the worst they will ever be.
Agents take the rowing while people steer. Agents handle more execution, and steering moves up layers of abstraction. A person still has to decide the direction. Seshan says the human brain stays valuable for expression, choosing what to build and caring for each other.
Product management is fast, empirical hypothesis testing. The core job is finding the most essential question about a product, testing it quickly with users and refining the hypothesis. Seshan values being prolific and empirical over being academic. Software resembles filmmaking, where a big budget does not guarantee a good result.
The best AI users expand their ambition. Seshan says strong users expand what they can do rather than just automating tasks. She cites Tyler Cowen's suggestion to ask whether work could be tried at 10X scale, and Patrick Collison's list of unreasonably fast, ambitious projects.
Ship early and iterate tightly with users. Seshan says her previous companies valued polish over shipping early. The Codex app improved through tight iteration with users, and outside attention came as the market caught up. Work mode runs on Codex's engine with some coding interface removed.
Writing as thinking stays, writing as reporting is automated. Seshan will not automate writing that helps her think, but models can handle writing that reports. She now communicates more through mocks, prototypes and experiment results than documents, and a former manager advised getting a doc to 70 percent before seeking buy-in.
Building with Sites and visualize in Codex
Sites in Codex builds a hosted site from a single prompt, so Seshan can make the exact tool she needs instead of configuring blocks in tools like Notion.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
The /visualize command builds visualizations from a user's past work, such as their ChatGPT usage.
Lenny builds a site about Seshan live during the episode.
Lessons from Stripe and Sutter Hill
At Stripe, strategy was a closely held treasure trove and briefs circulated widely. At OpenAI, which is founder-led with little top-down direction, strategy quickly becomes public.
Sutter Hill is an intentionally illegible firm that operates quietly, and Seshan credits it with an incubation playbook started by Mike Speiser.
Seshan says product marketing fit may come before product fit and can be tested by pitching about a hundred people before committing to the product's shape.