AI’s third era: the rise of persistent AI coworkers | Tara Seshan (Product Lead ChatGPT Work)

Tara Seshan frames the short history of AI products as three phases: chat, then agents, then a persistent coworker that gets work done alongside you. She is deliberately tentative about the third phase, calling it something that might come soon rather than something already shipped, and she describes the second phase as still largely a coding-agent phenomenon that has yet to reach most knowledge work.

that third era that might come soon is how do you work with a persistent coworker who is able to get things done with you

The coworker model is defined by cadence as much as by capability. Seshan describes it as the rhythm of working with a teammate: the agent does a batch of work, you give input, it works again, and you review work in progress at intervals. The appeal she names is that this is a more natural interface for people than issuing one-off commands.

Her evidence that today's setup is insufficient is drawn from inside OpenAI. Agent work is currently one-on-one, and colleagues used to paste screenshots of their Codex threads into Slack simply to show how they reached a number. She calls that workaround not quite the most natural way for people to collaborate, and asks what the right interface for shared agent work would be.

Lenny Rachitsky pushes the idea further, imagining his own agent checking another person's analysis for soundness. Seshan's answer is a design goal rather than a description of a shipped feature: work should be shared, with several people steering their agents together, and she closes the thread by saying these are things the team is thinking about.

Ideally, work feels like a multiplayer game

Asked twice what surprised her about OpenAI, Seshan separates the familiar from the unfamiliar. High growth, high talent and high intensity were known territory from earlier jobs, and so was the calibre of colleagues. What differed was how little direction arrives from the top.

She calls the company founders-led in a literal sense: every person is effectively a founder of their own area, with a very thin distance between the product team and the market, and full responsibility for finding product-market fit rather than inheriting a plan.

everyone inside the company, especially in their area, is in essence a founder to some extent

Her second surprise inverts her expectation. She arrived looking for the internal equivalent of a payments bible — settled doctrine she could learn from, as she had seen at earlier companies — and found instead that thinking about how the world or the product should look becomes public product or public messaging very quickly.

And actually OpenAI is open.

Rachitsky jokes about a secret room where AGI runs the master plan; Seshan answers that she is at least not in that room, then redirects to the observable claim: the cycle from internal idea to something users can touch is faster than anywhere she has worked.

That is testimony from one product leader about her own area, not a company-wide measurement. The episode offers no figures on release cadence, so the comparison with her prior employers rests on her judgment rather than on data.

Her account of what changes begins with what a slower market rewards. In an established market such as payments, a product leader can map competitor responses and reason from first principles through the next moves, and the market itself demands that rigour because most decisions were predictable in hindsight.

Winners will think more rigorously than everybody else.

In the current market she inverts the emphasis. The future is hard to read, the field is emergent and fast-changing, and staying tied to research matters more than writing a long reasoning document that resembles a PhD thesis. Being prolific and empirical, she argues, beats being academic and theoretical.

The switch was uncomfortable rather than obvious. She describes catching herself wondering whether she was skipping due diligence or being insufficiently thoughtful, and resolving that the thinking should instead be spent on defining the sharpest possible hypothesis — what she calls, borrowing Shishir Mehrotra's term, the eigenquestion — with everything else in a grand strategy treated as irrelevant.

What survives is the loop: find the essential question that determines whether the product works, design the fastest effective test, read the results, refine the hypothesis, run it again. She argues this was always the job and is now the most important thing in the company, with EMs, engineers, data scientists and designers all working the same loop.

The trappings have thinned. Running execution on time, writing documents and building presentations matter less in her telling. That is a claim about where emphasis sits inside one company, not a measurement of how PM time is actually spent, and Rachitsky frames it as the part of the role that persists while everything around it shifts.

The most concrete rule Seshan offers is a timing rule. Building against today's model and building against a model a year away are, in her account, equally wrong, and the only defensible horizon is a couple of months ahead.