从划桨到掌舵：OpenAI 如何把 Agent 做成持久共事的同事

Seshan has been at OpenAI for about a year. Much of it felt familiar from earlier high-growth companies, including talented colleagues and high urgency. What surprised her was that, unlike her past employers, OpenAI is not founder-led in the usual way. Top-down direction is very limited, and people feel like the founder of their own product area.

That keeps her close to the market, she says. Larger companies can insulate people from what users want, but at OpenAI you do whatever it takes to find product-market fit for your product, much as a founder would. She says she continues to feel like a founder, as she did in her earlier founding journey.

The second surprise is that she expected a treasure trove of secret strategy, like a company's internal payments bible. Instead, she says, OpenAI is open: views about how the world should look or how products should be built quickly become public product or messaging. What inspires her is that the company's work becomes something users can touch faster than anywhere else she has seen.

Seshan contrasts a static market like payments, where you can reason rigorously about competitors' next moves, with AI, where what emerges is hard to predict and staying tied to research matters most. In that setting, she says, being prolific and empirical is worth more than being theoretical. At her past companies she would write a long reasoning document, almost like a thesis. At OpenAI the question is how to get something in front of users as fast as possible.

That switch initially felt jarring, as if she were not doing her due diligence. She says the thinking you must do becomes pointed: define the core hypothesis, what she calls the "eigenquestion", the single most important thing to test. Any grand strategy beyond that is not relevant.

She argues that this has always been the heart of PM work, behind the trappings of running execution and writing documents. A PM pulls together understanding of users, the market and the technology to make the sharpest hypothesis, then makes the test fast. Now engineers, data scientists and designers are all thinking this way too, and she says many of the other trappings of the role have fallen away.

Asked about loops, in which an agent is told what success looks like and keeps going until it gets there, Seshan says work will increasingly resemble steering rather than rowing. Agents will do much of the rowing, and the steering will move to higher levels of abstraction, from a line of code to a goal. But she believes a person still has to decide the direction and give feedback.

Part of steering, she says, is what the data tells you, but a lot is an opinionated call. She says we underrate the power of intuition and of wanting the future to look a certain way. She expects people to steer groups of agents together, bringing teammates into the interaction between a person and an agent.

Lenny suggests that when everyone has the same tools, the human becomes the differentiator. Seshan likens this to fashion: clothes are functional, but what you wear makes a statement because it contrasts with what others do. She cites a Collison line that software is not like real estate but more like filmmaking, where money does not guarantee a good film. It takes opinion and artistry.

For the next three to six months, Seshan sees people working with agents at higher levels of abstraction and treating them as persistent teammates. You give input, they work, and you sync at different cadences, looking at each other's in-progress work. She says this coworker model feels like a more natural interface for agents, and she already sees it inside OpenAI.

The second shift is from one-on-one work with an agent to working together. Internally, she says, people used to share screenshots of their Codex threads in Slack, which is not a natural way to collaborate. She wants work to feel like a multiplayer game in which everyone steers their agents while agents handle more tactical tasks.

She stresses that making this possible is not only about intelligence. Local agents are convenient because they can see the data on your machine, while cloud agents need substantial infrastructure. An agent also needs access to third-party systems, like a hire who is locked in a room without access to Google Docs, Slack or the company database. She calls these data access, cloud and reliability pieces prosaic but just as important.

Lenny argues that when easy things become easy, what separates people and companies is how ambitious they are. Seshan says the most effective people use AI not just to automate rote work but to expand what they can do. Where the unicorn used to be a product-minded engineer-designer, she says everyone can now spin up designs, build a prototype and model pricing scenarios.

She says the hardest part is expanding your thinking about what is possible, and she points to Patrick Collison's list of unreasonably ambitious projects done quickly. All of them predate today's tools, so she expects many more now. Citing Tyler Cowen, she says a big PM contribution is to ask others whether they could try something faster or at 10x scale.

She shares two internal memes. "Is this maximally accelerated?" came from Nick Turley, and "are you mainlining it yet?" is a way of asking whether you use the product all day. She adds that "feeling the AGI" is part of the culture, and that the team must build for where models will be in two to three months. Building for last year's model, or one a year out, are both wrong, and she says tight communication with research helps the team judge the timing.

Lenny asks about the toggle between Chat and Work in ChatGPT. Seshan says the north star is that users do not have to choose: you type your task and the system picks the right harness and model. Meanwhile, she says Codex users should keep using Codex, and ChatGPT users who want the new agent capabilities should use Work mode, which runs on Codex under the hood.

The difference between Work and Codex, she says, is really at the UI level: how much technical detail and chain of thought you see. Work removes some of the coding interface but keeps the same power, and she says OpenAI's corporate finance team uses it to build complex financial models. The aim is to meet people where they are while bringing a billion ChatGPT users the power of agents.

Seshan frames this as a move from chat to agents and possibly to persistent coworkers, and says the challenge is making it easy without users needing to understand harnesses. She contrasts this with her past companies, where polish was king. In this era, she says, getting a transformative product into users' hands is far better than perfect, and "done is better than perfect".

Asked what changed to make Codex take off on Twitter, Seshan says little changed internally. The team that built the app was user-focused, iterated quickly and used the product constantly, and the market caught up. She says the same questions apply: whether the mission is elevated enough, whether progress is maximally accelerated, and whether they are mainlining it. Lenny notes that it is the humans, not AI, who made the difference.

On blurring roles, Seshan says she likes startups with few boundaries, as at Stripe, where anyone could do anything. What she wants is someone accountable for whether the product is used, wanted and high quality, whether that person is an engineer, a designer or a PM. She admits she does not know how to keep honing craft when models abstract some of it away.

Asked where human brains will stay valuable, she names accountability, since someone must own the outcome and the agent is like a report. She adds expression, as in the film analogy, and caring for and relating to one another, such as how a team elevates each other's ambitions.

Seshan says she now builds sites all the time. She built one for her team to play a game, and one for a backpacking trip that tracked the route's elevation and let everyone enter food. She sees sites as a realization of the old dream of malleable personal software, because you just prompt for the exact tool you need, and they can be shared, hosted and updated with internal data.

A site is also a more dynamic way to present than a slide deck, and she says it has replaced many of the docs and sheets she once produced. You can ask for one in Codex, Work, on the web or on mobile. Lenny tries it live by asking for a site about her.

Her other tip is the slash command `/visualize` in Codex, which can build a visualization of something like your ChatGPT usage. She says it makes presenting charts and data in a way that tells a story simple and surprisingly delightful.

Seshan distinguishes writing as thinking from writing as reporting. Reporting, like a weekly status or a launch plan, she happily automates with models. Writing as thinking, such as a brief arguing for a strategy, she says she never will, because outlining, cutting and editing is how she gets her ideas in line. She thinks both extremes of always or never using models for writing are wrong.

Her process, honed at Stripe, is to write a brief on a new idea, then have people attack it and poke holes, and take it to the next person. A manager once told her to write to 70% completion and then get buy-in to reach 100%, because people do not engage with a perfectly polished idea. At OpenAI she still writes constantly but for herself.

What she shares now is mocks, prototypes or A/B results rather than long documents, since a long doc is no longer a signal that you thought something through. To avoid atrophying her thinking, she follows a rule: if she makes people read a document or sit in a meeting, she should have invested at least that much time. She starts and ends the writing herself and uses AI in the middle for research, data and pushback.

Lenny asks about her stint as an entrepreneur in residence at Sutter Hill Ventures. Seshan describes it as an intentionally illegible firm that operates quietly yet is behind iconic successes, using an incubation model started by Mike Spiser. She says it showed her that finding product-market fit and building very large companies is not just luck, and that the firm has a repeatable playbook.

She says her career has been about finding product-market fit repeatedly, and that Sutter Hill has worked out how to do it for B2B products. The firm's recruiting is unparalleled, and she mentions an internal tool called Reticle that maps people they know and the ten best people those people know.

Her biggest lesson was that she had underrated product marketing fit. The narrative and positioning can come before building: pitch a hundred people, refine the pitch, and only then commit to the product shape. She says she once saw product marketing as glue between functions and now sees how transformative it can be.

Seshan says knowledge work differs fundamentally from coding. With code you can verify the output with tests or by trying it. With a deck of numbers, you cannot simply see the result and trust it, and you have to think about the process, inputs and reasoning.

That shapes the product: ChatGPT should be a collaborator, showing in-progress work, citations and inputs, so you can go on the journey with the model and know the result is right. She asks whether a thread, so suited to coding, is the right surface for this, and whether chain of thought should show more citations. Lenny likens it to proving the work to an executive audience.

Seshan recommends *Barbarian Days* by William Finnegan, which she reads as a story about being passionate and dedicated without being good at something, and *Anna Karenina*, which she says reads differently at 13, 17 and 30. She says that is like the growth the new era will demand. Of films, she mentions *The Odyssey*, which she reads as being about AI, and *Rashomon*, which she says shows what creativity can do under constraints.

Her favorite AI products outside OpenAI are the ones her friends build for her, part of what she calls the cozy software movement. She names a friend's app that turns anything into a podcast in your podcast app, and a private social network for friends. Her life motto is Toni Morrison's four lines, including that you make the job and it doesn't make you, and that you are not the work you do.

She calls the Thiel Fellowship an inflection point because someone elevated her ambitions. As a closing tip she urges people to try ChatGPT's cloud work on mobile: start a task, go offline on a subway ride, and find it finished when you resurface.