OpenAI product lead Tara Seshan says the company's third era of AI is here: persistent coworkers that work alongside people over time, not just chatbots or one-off agents. She described this shift on Lenny's Podcast, where she leads Codex and ChatGPT Work, OpenAI's fastest-growing product for knowledge workers.
Seshan frames era one as chat and era two as agents working with users. Era three is a persistent coworker that gets work done with you, collaboratively and over time. That means agents that feel like teammates, with shared context, longer loops, and joint steering across people.
Seshan spent six years at Stripe as one of its first five product managers. She later led product at Watershed, founded a company, and is a Thiel Fellow.
Building for a three-month horizon
She said you fail if you build for where models are now and you fail if you build for where they will be in a year. The only way to build is for two to three months out, tight to research, with the model as the center of the product. That tight coupling shows in how she describes roadmap alignment with researchers on specific capabilities like coding and writing.
In slower markets, rigorous strategy docs predicted competitor moves and won. In this market, being prolific and empirical beats being academic or theoretical. The core PM job narrows to sharp hypothesis definition, fast tests with users, and feeding results back into the loop.
Seshan describes future work as steering, not rowing, where agents do the rowing and people set direction. Abstraction rises from tab completion to goal-level direction and then higher, but humans still make the opinionated call. That opinion is the differentiator when everyone has the same tools, much like authorship in film.
Three modes, one box
Most work today is one human with one agent and spawned subagents. Internally, teams were sharing Codex thread screenshots on Slack to prove how a number was reached. The next interface is multiplayer, where groups of humans steer groups of agents together with shared data access.
Seshan laid out three modes today: Chat mode for search and conversation, Work mode for knowledge work, and Codex mode for development. Work mode is Codex under the hood with the coding UI removed, so worktrees and chain-of-thought detail are hidden. The north star is no toggles at all, where the system picks the right harness and model from a single prompt.
For a billion monthly ChatGPT users, OpenAI chose to ship Work inside chat on web and desktop rather than wait for polish. Seshan said done is better than perfect when conviction is high that agents transform work, and iteration after launch beats waiting. Decomplexifying harnesses and model choices matters more to adoption than pixel perfection at this stage.
What actually drives quality
Seshan credited no strategy change for Codex momentum, only the same loop: dogfood, or mainline, the app all day and listen. The team that built the desktop experience fixed their own pain and shipped, and external perception caught up later. She framed that as humans, not models, driving quality through obsession with user signals.
Intelligence gains matter, but Seshan stressed prosaic blockers like local data access and cloud infrastructure. An isolated cloud agent without Google Docs, Slack, and databases is as limited as a new hire locked in a room. Reliability and system access determine end effectiveness as much as longer context or stronger reasoning.
With rote tasks automated, ambition becomes the scarce input. Seshan said the best users expand what is in range and then realize more of what is in their heads, like the old unicorn PM who could code, design, and price. PM craft now includes elevating others' ambition and reminding teams the ceiling is meaningfully higher.
Three phrases shape shipping: is this maximally accelerated, are you mainlining it yet, and feeling the AGI. Mainlining replaces dogfooding and means using the product to do real work all day and bringing taste to the feedback. Feeling the AGI keeps the mission present in product decisions without pretending a secret master plan exists in a locked room.
Seshan expected a treasure trove of secret strategy on arrival and found the opposite. OpenAI is founder-like, with very limited top-down direction and thin distance between product teams and the market. That founder distance to users speeds the cycle where ideas become public product or public messaging within weeks.
Why this matters for product development
If work becomes steering fleets of persistent agents, the interface winners are those that solve collaboration, permissions, and provenance across many agents at once. Incumbents with distribution like ChatGPT can push agents to a billion users in one move, while developer-led tools must win on depth and trust with engineers. The two-to-three-month horizon also reorders roadmap bets, favoring teams that can ship, measure, and rewrite quickly over those that bet on a 12-month plan.
For product teams, the practical takeaway is to build for a short horizon with the model at the center, and to treat daily internal use as the primary feedback loop. For those developing their own skills in this area, an AI Learning Path for Product Managers can help ground these concepts in hands-on practice.
Watch whether Work collapses into ChatGPT as a single box that chooses harness and model automatically. Watch for multiplayer agent primitives, shared memory, and enterprise connectors that make cloud agents truly useful for finance, operations, and research teams. If those ship before competitors normalize them, the third era stops being a demo and starts being how teams actually work.
Your membership also unlocks:
