AI's third era is the persistent coworker, and the product cycle is only three months
OpenAI product lead Tara Seshan argues that AI products are moving from chat, through agents, toward persistent coworkers, and that you have to build for the model capabilities of two to three months out; future work is steering rather than rowing, and writing-as-thinking should never be automated.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Build for the model capabilities of two to three months from now
Tara divides the evolution of AI products into three eras: the first was chat, the second was collaborating with agents, and the third is arriving now — working alongside a persistent coworker that gets things done with you. She stresses that when you build a product, you should base it neither on today's model capabilities nor on your prediction of what models will do a year from now, but on the model capabilities of two to three months out, because both extremes fail.
— Tara SeshanThe work of the future is steering, not rowing
Tara believes future work will be much more about ‘steering’ than ‘rowing’: agents handle execution, humans set direction. The level of abstraction at which you steer will keep rising, but in the end humans still have to make opinionated decisions. She stresses that even as the altitude of steering goes up, human judgment and accountability at the critical junctures remain irreplaceable — that is both the responsibility and the value.
— Tara SeshanThe endpoint of the product is that users never pick a tool
Tara explains that ChatGPT's ‘work mode’ is Codex underneath, with a different UI. The product north star is that users should not have to choose — the system picks the right tool for the task on its own. She stresses that ‘done is better than perfect’: get it into users' hands first, then correct based on feedback; shipping fast matters more than polishing. And she says this is the worst models have ever been.
— Tara SeshanCodex did not change; users' perception did
The host observes that the conversation on Twitter about Claude Code and Codex has shifted over the past few months, with more people leaning toward Codex. Tara responds that the Codex team has stayed user-oriented and fast-iterating all along, and nothing about how it operates internally has changed — the market and users have simply come to realize it. In her view the change is not in the team but in users' perception.
— TaraThe writing-as-thinking part should never be automated
Tara distinguishes ‘writing as thinking’ from ‘writing as reporting’. The former she absolutely will not automate, because it is how she sorts out her thinking; the latter she hands off entirely to the model. She recommends taking a document to 70% and then getting feedback from someone, rather than chasing perfection. At OpenAI she still writes a great many documents, but mostly for herself, because a long document is no longer a signal of depth of thought.
— TaraBefore product market fit comes product marketing fit
The most important lesson Tara learned at Sutter Hill: product market fit matters, of course, but she had underestimated ‘product marketing fit’. Before you build the product, you should sharpen the narrative and positioning by pitching it a great deal. She believes outstanding product marketing work can decide whether a company lives or dies, and notes that Mike Speiser is unmatched at it.
— TaraIn their own words · checked verbatim
You fail if you build for where the models are now. You fail if you build for where you think the models will be in a year. Like both outcomes are equally wrong.
Tara Seshan27:14
This is the worst the models will ever be.
Tara Seshan29:14
I think humans will continue to be most valuable as a certainly as a um like an entity of accountability.
Tara46:25
writing is thinking is something I never will automate.
Tara53:30
product market fit is sure important but actually I really underrated product marketing fit.
Tara1:02:43
the teal fellowship was an inflection point in my life I wouldn't be where I am without it
Tara1:16:51
you've finally got these things running in the cloud doing real work
Tara1:20:56
Figures
| Model capability planning horizon | 2-3 months | 27:14 |
| Codex monthly active users | 10 million | 33:16 |
| Recommended document completeness | 70% | 57:36 |
| Teal Fellowship recipients per year | 20 | 1:17:52 |
| Teal Fellowship grant | $100,000 | 1:17:52 |
Glossary
- persistent co-worker
- An AI agent that works with you over the long run and takes on tasks proactively, as distinct from one-off chat or short-lived agents.
- product marketing fit
- How well a product's narrative and positioning match the needs of a market segment, with the emphasis on sharpening positioning early through pitching.
How to listen
AI product managers, founders, and investors, plus engineers who want to know what work will look like.
You can skip the discussion of craftsmanship at 42:22 without losing the through-line.