Lenny's PodcastAI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
CHAPTERS
- 0:00 – 2:40
From chat to agents to “persistent AI coworkers” (the three eras framework)
Tara frames AI product evolution as three eras: chat interfaces, agentic tools, and an emerging phase where AI becomes a persistent coworker. Lenny and Tara discuss the gap between what AI can do and how most people are using it today.
- •Three eras: chat → agents → persistent AI coworker
- •“Overhang” between AI capability and real-world usage
- •Why the next interface may feel like working with a teammate, not a tool
- 2:40 – 6:34
What’s surprising about OpenAI: “founders-led,” thin distance to users, and radical openness
Tara shares what most surprised her about OpenAI’s culture: limited top-down direction, strong founder-like ownership across teams, and fast cycles from idea to user-visible product. She also dispels the notion of a hidden master strategy, describing OpenAI as unusually open with its thinking.
- •Founder-like ownership across many people and teams
- •Very thin insulation between product teams and real user needs
- •No “secret strategy room”—ideas quickly become public-facing product/messages
- •Speed of shipping and feedback loops as a defining cultural trait
- 6:34 – 10:57
Why AI product strategy is ‘empirical’ now: hypothesis-first, short cycles, and the “Eigen question”
In fast-moving AI markets, Tara argues that grand long-range strategy breaks down; the core PM skill becomes defining the sharpest hypothesis and testing it quickly. She contrasts theoretical rigor in stable markets (like payments) with rapid experimentation required in frontier AI.
- •Strategy becomes less predictive; execution becomes test-driven
- •Prolific experimentation beats long reasoning docs
- •Define the core hypothesis (“Eigen question”) and test it fast
- •PM craft centers on loops: hypothesize → test → learn → iterate
- 10:57 – 15:36
The shift from rowing to steering: loops, abstraction layers, and opinionated direction-setting
They explore how agents expand “loops” beyond engineering into product and knowledge work. Tara describes a future where humans steer at higher abstraction layers while agents do the rowing—yet humans still provide judgment, taste, and directional intent.
- •Future work: humans steer, agents row
- •Steering moves up layers of abstraction over time
- •Opinionated calls and intuition remain differentiators
- •Teams may steer fleets of agents collaboratively
- 15:36 – 20:05
Agents as teammates: persistence, multiplayer collaboration, and the infrastructure reality
Tara predicts a move from one-on-one agent usage to persistent coworker-like agents, and then toward multi-user collaboration with agents. She emphasizes the unglamorous enablers: data access, cloud infrastructure, integrations, and reliability.
- •Persistent agents that behave like teammates/coworkers
- •Need for collaborative/multiplayer workflows with agents
- •Sharing threads/screenshots is a temporary collaboration workaround
- •Infra matters: access to docs, Slack, systems; cloud agent reliability
- 20:05 – 26:44
Ambition as a PM superpower: widening the ‘range of possibilities’ and raising the ceiling for teams
They discuss why ambition becomes the key differentiator when AI makes many tasks easier. Tara explains that effective AI users expand what they can personally execute (design, prototype, modeling) and that PMs increasingly elevate others’ ambition and speed.
- •AI turns more people into ‘unicorns’ across disciplines
- •Biggest limiter becomes imagination, not execution capacity
- •PM role: elevate ambition, push for 10x scope and faster timelines
- •Internal memes: ‘maximally accelerated’ and ‘mainlining it’
- 26:44 – 29:21
Building for models that don’t exist yet: the 2–3 month rule and tight research alignment
Tara explains why you ‘fail’ if you build for current model capabilities or for a speculative 1-year future—both are wrong. The practical approach is to build for the expected state 2–3 months out, staying tightly coupled to research priorities and model trajectories.
- •“Worst the models will ever be” mindset
- •Avoid building too tightly around today’s limitations
- •Avoid overfitting to a distant, uncertain future
- •Stay aligned with research roadmaps and near-term capability targets
- 29:21 – 33:59
ChatGPT vs Codex, and Chat vs Work: what the modes mean (and the ‘no toggles’ north star)
Lenny asks Tara to explain the app’s current complexity: product selection (ChatGPT vs Codex) and mode selection (Chat vs Work). Tara describes the intention: meet users where they are today while moving toward a future where users simply state tasks and the system picks the right ‘harness.’
- •North star: users shouldn’t choose models/modes—system routes tasks automatically
- •Codex: dev-oriented UI; Work mode: Codex power with a knowledge-work-friendly UI
- •Chat mode: conversation/search; Work mode: “get things done” agentic flow
- •Goal: bring agent power to the massive ChatGPT user base without complexity
- 33:59 – 38:56
Shipping quickly at massive scale: done > perfect, iterate on real signals
Tara contrasts earlier product eras where polish dominated with today’s urgency to get transformative capabilities into users’ hands quickly. They discuss how confusion post-launch is expected, and the real differentiator is rapid iteration guided by strong signals.
- •Polish used to be king; now speed and learning are paramount
- •Ship early when conviction is high; refine quickly after
- •Pre- and post-launch iteration loops both matter
- •Listening to the right signals is the core operating system
- 38:56 – 42:20
The Codex ‘vibe shift’: why momentum changed (hint: it’s the team’s operating model)
Lenny asks about the market’s shift toward Codex. Tara credits consistent internal behavior—dogfooding/mainlining, fast feedback loops, and founder-level ownership—more than any single tactical change, with the market eventually noticing.
- •Codex team focus: obsessive usage, rapid fixes, tight loops
- •External perception lagged behind internal progress
- •Founder-like independence: people notice issues and just build
- •Human obsession and craft drive the product edge
- 42:20 – 45:54
Roles blur in the agent era: boundaryless teams vs preserving craft
They discuss how AI tools erode traditional boundaries between PM, design, and engineering, returning teams to startup-like ‘everything and nothing is your responsibility.’ Tara also reflects on the tension between fluid collaboration and maintaining deep craft in a discipline.
- •AI enables broader execution across roles (PM/eng/design overlap)
- •Accountability still matters: someone must be DRI for outcomes
- •Fluid roles feel like startups (Stripe-style boundaryless work)
- •Open question: how to preserve/advance craft as models abstract tasks
- 45:54 – 48:18
Where humans keep unique value: accountability, expression, and caring for each other
Tara identifies enduring areas of human advantage: owning outcomes, creative authorship and taste, and the social/emotional work of motivating teams. She argues these become even more important as agents take on more tactical work.
- •Humans as the locus of accountability and responsibility
- •Taste, authorship, and expression in product/software
- •Relationships: collaboration, ambition-setting, team energy
- •Regulated and high-stakes contexts reinforce human oversight
- 48:18 – 52:34
How Tara uses AI day-to-day: Sites everywhere and the magic of /visualize
Tara shares concrete personal workflows: building ‘Sites’ as fast, shareable software artifacts instead of static docs, and using /visualize to turn data into compelling charts and stories. Lenny demos generating a site about Tara live, highlighting how quickly presentational artifacts can be created and iterated.
- •Sites as ‘malleable personal software’ (dashboards, games, trip planners)
- •Sites replace docs/slides as dynamic, shareable artifacts
- •/visualize turns usage/data into instant, story-friendly visualizations
- •Shift from manual artifact-building to prompt-driven creation
- 52:34 – 1:00:16
Writing to think vs writing to report: staying sharp and avoiding AI ‘brain rot’
Tara explains her distinction between writing as thinking (never automated) and writing as reporting (heavily automated). She shares tactics for producing strong briefs—write to 70%, then iterate with stakeholder feedback—and a personal rule for respecting others’ time.
- •Don’t automate ‘writing as thinking’; do automate ‘writing as reporting’
- •Docs aren’t proof-of-thought anymore; prototypes/results communicate better
- •Write to ~70% then co-polish with key stakeholders
- •Rule: if you make N people spend X time, you should prep at least that much yourself
- 1:00:16 – 1:21:44
Lessons from Sutter Hill: product marketing fit before product market fit + lightning round closing
Tara reflects on Sutter Hill’s incubation playbook and her biggest takeaway: narrative/positioning can be the first thing to test, even before building. The episode closes with a lightning round on books, films, favorite AI tools, life mottos, and her Thiel Fellowship experience—ending with a call to try ChatGPT Work and agentic workflows across devices.
- •Sutter Hill’s repeatable playbook for building iconic B2B companies
- •Underrated insight: product marketing fit (positioning) can precede building
- •Pitch/refine with many customers before committing to product shape
- •Lightning round: book/film recs, cozy software, Toni Morrison motto, Thiel Fellowship story, and trying ChatGPT Work
