AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)

What It Covers

Tara Seshan, OpenAI's product lead for ChatGPT and Codex, outlines how AI product development is entering a third era—moving from chat to agents to persistent AI coworkers—while explaining how the PM role must shift from theoretical strategy documents toward rapid empirical testing cycles with two-to-three month model capability horizons as the planning unit.

Key Questions Answered

• •AI Product Planning Horizon: Build for where models will be in two to three months—not current capabilities, not twelve months out. Both extremes produce equally wrong outcomes. Current capabilities make you too conservative; twelve-month projections make you too speculative. Staying tightly connected to research roadmaps and specific capability improvement areas is the only reliable way to calibrate this window accurately.
• •Empirical Over Theoretical PM Work: The core PM skill that survives AI disruption is hypothesis sharpening—identifying the single most essential question (Shishir Mehrotra's "eigen question") and testing it as fast as possible. Replace lengthy strategy documents with prototypes users can touch. The shift from academic reasoning docs to rapid testable artifacts is the biggest operational change Seshan made moving from Stripe to OpenAI.
• •Writing as Thinking vs. Reporting: Separate writing into two distinct modes. Writing-as-thinking—briefs, strategy, product rationale—should never be delegated to AI because the act of drafting, cutting, and iterating is itself the thinking process. Writing-as-reporting—status updates, launch plans, summaries—should be fully automated. Share briefs at 70% completion to invite collaborators to polish rough edges rather than deflect off finished ideas.
• •Ambition Elevation as Core PM Responsibility: The PM role now includes actively raising teammates' ambition ceilings. Because AI tools expand individual capability dramatically, the limiting factor shifts from execution capacity to imagination. Referencing Patrick Collison's "fast" page—a list of unreasonably ambitious projects completed quickly before modern AI—Seshan argues those benchmarks should now be considered the floor, not the ceiling, for what teams attempt.
• •Knowledge Work vs. Coding Agent UX: Coding agents succeed because outputs are verifiable through tests and runtime behavior. Knowledge work agents require a fundamentally different product approach: users need visibility into citations, reasoning chains, and in-progress steps to trust the final output. This means ChatGPT's work mode must surface the collaborative process, not just deliver a finished artifact, mirroring how an executive presentation requires showing the analytical work behind conclusions.
• •Product Marketing Fit Precedes Product Market Fit: From her time as EIR at Sutter Hill Ventures, Seshan learned that the narrative and positioning of a B2B product should be tested before committing to a product shape. The process: pitch one hundred potential customers, refine the story until it lands consistently, then lock in the product architecture. Mike Speiser's repeatable incubation success at Sutter Hill—producing companies including Snowflake—is built on this sequencing discipline.