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

I think AI product evolution goes from chat, to agents that carry out work, and then to a persistent coworker that keeps making progress with you and your team. Models move so quickly that designing only for today is wrong, but betting on a year from now can be wrong too. Two or three months ahead is the useful horizon.

There’s a huge gap between what AI can do and what people ask it to do. As a PM, what has had to change most?

I’ve become much more empirical. Rather than make a grand strategy argument, I want the sharpest hypothesis, a real thing in front of users, and fast learning. The rigor is choosing the one question that determines whether the product works, then testing it quickly. The PM job is still understanding users, market, and technology; making a precise bet; inspecting what happens; and running the loop again.

As agents take longer-running tasks, work looks less like rowing and more like steering.

Exactly. Someone still has to set direction, react to new information, and make an opinionated call about the future they want. Agents can do more rowing; people will steer together. Software is more like filmmaking than real estate: spending more doesn’t guarantee a good outcome. Products need authorship, taste, and something interesting to say.

You joined OpenAI expecting talent, urgency, and intensity. What surprised you?

The surprise was how founder-led it feels. There’s little top-down direction, so you feel close to the market and accountable for product-market fit in your area. I expected a secret master strategy, but OpenAI is unusually open: ideas become product or messaging quickly. The cycle from work to something users can touch is extraordinarily fast.

Internally, we ask: are we ambitious enough, are we moving at maximum speed, and are you “mainlining” the product—using it constantly enough to bring your taste to every rough edge? That’s the new dogfooding. The mission stays present: AGI should be beneficial, while we stay connected to research and the coming capability curve.

You’ve argued AI should make us more ambitious, not merely more efficient.

The strongest users don’t just automate rote work; they expand what they can attempt. More people can sketch designs, build prototypes, model pricing, or test scenarios themselves. The limiting factor becomes imagination: can you recognize what’s possible on an unreasonable timeline? We should expect far more unusually ambitious projects.

For people seeing ChatGPT chat, work, and Codex, what are they choosing?

Our north star is that they shouldn’t have to choose. Describe the task, and the system should select the model and harness. For now, Codex is development-oriented, Chat is conversation and search, and work gives ChatGPT users agentic capability through a less technical interface. We want knowledge workers to access that power without learning developer concepts.

Intelligence alone isn’t enough. An agent without access to your data and systems is like hiring a colleague and refusing them your docs, Slack, or database. Cloud agents need permissions, reliability, and third-party connections. And knowledge-work agents must bring you along: for a strategic deck, I need sources, intermediate work, reasoning, and context—not just an answer.

You still write a lot. How do you use AI without letting it replace your thinking?

I separate writing for thinking from writing for reporting. I automate status updates, launch summaries, and format conversion. But for strategy or a spicy argument, I write myself; outlining, drafting, cutting, and revising get my ideas straight. I’ll take a brief to seventy percent, then invite the people whose buy-in I need to improve it. Increasingly, though, the best artifact is a mock, prototype, or experiment result—not a long document.

My guardrail: if I expect someone else to read a document, I should have spent at least that much time making it. AI can research, retrieve data, or challenge an assumption, but I don’t want it generating the first thought or sanding away my prose. I build small sites, dashboards, and shareable presentations all the time now; Codex makes turning information into a legible story delightfully easy.

Humans remain valuable through accountability, expression, and care for one another. Somebody has to own whether the outcome is good, safe, useful, and appropriate. Teams still need to build enthusiasm, learn together, and raise each other’s ambitions. A fellowship early in my life told me my ambition could be larger. Use these tools to make more possible, but keep your judgment, authorship, and responsibility in the loop.