Aug 30, 2026 · 1h 22m
Persistent AI coworkers redraw the boundaries of knowledge work
AI’s third era: the rise of persistent AI coworkers | Tara Seshan (Product Lead ChatGPT Work)
As AI shifts from answering prompts to carrying work forward, people must decide what to delegate, what to verify, and what thinking to protect.
- 1Product managers should replace fixed grand strategies with empirical questions, rapid tests, and continuous refinement.
- 2Persistent agents may absorb tactical work, while humans retain accountability, judgment, context, and responsibility for outcomes.
- 3AI can accelerate reporting and execution, but important thinking still requires human ownership from first draft through final judgment.
Don't miss
Seshan explains why she starts and ends important thinking documents herself, using AI in the middle without surrendering ownership of the ideas.
The brief
Tara Seshan and Lenny Rachitsky examine AI’s move from chatbots and task-based agents toward persistent coworkers, arguing that product builders must learn by testing rather than planning for a fixed future.
As agents take on tactical work, humans may steer at higher levels—but the delegation only works when people retain accountability, judgment, context, and responsibility for outcomes.
The conversation treats ambition as a product advantage: when routine execution becomes easier, the differentiator is the scale of the problems teams choose to attempt.
Seshan’s practical rule for avoiding AI brain rot is to begin and end important thinking documents herself, using AI in the middle for research, data gathering, and critique.
The episode’s sharpest distinction is between coding and knowledge work: tests can validate code, but knowledge agents need visible proof so people can assess how conclusions were reached.
What was said on this episode
47 statements · 39 positive · 3 negative · 2 mixed · 3 neutral
Persistent AI coworkers may soon become a third era of AI products.
“that third era that might come soon is how do you work with a persistent co-worker who is able to get things done with you?”
Listen at 0:07
Products fail when designed for current or incorrectly anticipated model capabilities.
“You fail if you build for where the models are now. You fail if you build for where you think the models are.”
Listen at 0:24
Empirical, rapid experimentation matters more than theoretical product planning.
“Being prolific and empirical is way more important than being academic or theoretical.”
Listen at 0:38
OpenAI moves research into user-facing products faster than other organizations.
“that cycle is faster than anywhere else I've seen.”
Listen at 5:29
AI product development should remain closely connected to research.
“most importantly, it's like very, very important to stay tied to the research.”
Listen at 7:44
Product management centers on identifying and testing the essential product question.
“the core of it has always been about what is like the most essential question you need to ask about your product?”
Listen at 9:29
AI agents will perform more tactical work while humans increasingly steer.
“the future of work will look more like steering than rowing”
Listen at 11:24
Human direction of AI systems will shift toward higher abstraction levels.
“the steering will continue to. maybe go up layers of abstraction.”
Listen at 11:58
Human judgment will remain necessary for choosing directions and goals.
“that is required from a person.”
Listen at 12:48
Human-agent product development loops are accelerating.
“those loops are moving faster and faster and faster.”
Listen at 15:21
AI agents are increasingly being designed as persistent teammates or coworkers.
“People are increasingly thinking about agents that are persistent, that feel like teammates, that feel like co-workers”
Listen at 16:26
Future AI-enabled work may resemble multiplayer collaboration among people and agents.
“Ideally, work feels like a multiplayer game where all of us together are getting stuff done”
Listen at 18:03
Cloud agents are ineffective without access to users’ systems and data.
“a cloud agent that is similarly isolated will not be that effective.”
Listen at 19:35
Effective AI users expand their capabilities rather than merely automate routine tasks.
“the people that we see who are most effective at using AI tools don't simply use it to automate rote tasks, but use it to expand the set of things that they are capable of doing.”
Listen at 20:42
AI tools have dramatically expanded what individuals can accomplish.
“the set of possibilities have widened dramatically.”
Listen at 22:04
Product managers should raise colleagues’ ambitions and awareness of AI possibilities.
“elevating others' ambitions or reminding them of what's possible here is a huge part of the product management role.”
Listen at 24:30
Ambition, acceleration, and intensive product use should guide product development.
“Those to me are like the three memes of product development that we just have to spread as much as possible now.”
Listen at 26:02
AI products should target model capabilities expected two to three months ahead.
“The only way to build is two to three months.”
Listen at 27:46
Product development should closely follow research agendas and roadmaps.
“ensuring that product development is as tied as possible to what research has as its agenda and its roadmap is really important.”
Listen at 28:45
ChatGPT should eventually choose the appropriate model and harness automatically.
“Our north star here is that users do not need to make decisions between picking between all these different options.”
Listen at 29:50
ChatGPT will select the appropriate model for a user’s task.
“It'll pick the right model for you to be able to get that thing done.”
Listen at 30:14
ChatGPT Work mode uses Codex underneath its interface.
“work mode, that's where under the covers, this is codex.”
Listen at 31:14
Codex can perform financial modeling tasks as well as ChatGPT Work mode.
“work mode and codex mode, if you go to codex and ask it to generate an amazing financial model to price your product or something like that, or like, tell me, predict my revenue for the next six months or something like that, codex will do as good a job as work mode.”
Listen at 32:26
OpenAI plans to extend agent capabilities beyond coding into knowledge work.
“We'd like to bring it to more more domains, certainly, like knowledge work.”
Listen at 35:04
Persistent AI coworkers may soon define a third era of AI products.
“that third era that that might come soon is how do you work with a like persistent co-worker who is able to get things done with you”
Listen at 35:37
Early release of transformative AI products is better than waiting for perfection.
“Getting the product in the hands of users when you have so much conviction that, hey, it's transformative, like is way better than perfect.”
Listen at 36:50
AI teams should iterate rapidly and respond to meaningful user signals.
“iterating as quickly as possible and listening to the right signals is, regardless of whether that's pre-launch, post-launch, ideally pre-launch, is the key thing.”
Listen at 37:44
Codex’s team improved the product through intensive internal use and rapid iteration.
“The team who initially got it up and running and were working on it were super, again, user-focused, tight iteration loop, really dog-fooded the thing, like mainlined the app as much as possible to get everything right.”
Listen at 40:07
Teams need clear accountability for product adoption, quality, and effectiveness.
“someone needs to look after the or have core accountability for is this product being used by users? Is it something that people want? Is it high quality? Is it effective?”
Listen at 43:29
AI models are abstracting some professional tasks and may perform them better than individuals.
“some pieces of our craft are actually getting abstracted by models being able to do it really effectively, maybe better than individuals can.”
Listen at 44:45
People will retain accountability for AI-generated outcomes, at least temporarily.
“Ultimately, who owns? What was the end product? Was it high quality? Was it the thing that you wanted it to do and say? That will certainly remain a person, at least for now.”
Listen at 46:35
Human creativity and expression will remain valuable in AI-assisted work.
“The human brain is also really valuable for expression.”
Listen at 46:56
OpenAI Sites enable flexible personal software through natural-language prompts.
“Sites kind of realized the dream of like malleable personal software”
Listen at 49:38
Codex can create a functional site from a natural-language request.
“In Codex, be like, create a site that is a, I don't know, is a mafia game for my team. And it will just do it.”
Listen at 50:32
Easy AI site creation has changed Tara’s daily work practices.
“The easy reach of building a site all the time has changed what my day-to-day looks like”
Listen at 51:00
Codex Visualize simplifies presenting charts and data clearly.
“Visualize makes that incredibly simple.”
Listen at 52:19
AI models should automate routine reporting work.
“Writing as reporting, I happily automate, or I use the models all the time to make that as simple as it can be.”
Listen at 53:42
Human-authored writing should remain central to developing ideas.
“writing as thinking is something I never will automate.”
Listen at 53:50
Routine reporting can be outsourced, but thinking-oriented writing should remain human.
“I will, again, outsource all writing is reporting as much as possible to the model, but writing is thinking I have to do myself.”
Listen at 58:39
Sutter Hill has a repeatable playbook for achieving product-market fit.
“there is there is like clearly a way to do it. There's clearly a roadmap for making that possible. There is a set of things one can do to get this repeatedly. It's not just luck. It's not just a dark art.”
Listen at 1:01:39
Product positioning should be tested before building the product experience.
“product marketing fit, that narrative, that positioning is actually even before you build a product experience, the right thing to test.”
Listen at 1:03:28
Coding agents can often be validated through tests of their outputs.
“coding is so output-oriented that when you ask it to do a coding task, you can verify whether it did the task correctly or well via tests.”
Listen at 1:05:21
Knowledge-work outputs cannot be reliably validated by inspecting final artifacts alone.
“knowledge work is different in that I can't simply look at the deck in the end and see the numbers.”
Listen at 1:05:37
ChatGPT Work should expose process, inputs, and citations for knowledge-work tasks.
“a lot of work. that we have done and have to continue to do is continue to adapt the product to knowledge work, which means way more focus on making ChatGPT your collaborator”
Listen at 1:05:56
Small, friend-built software tools are an appealing future of software.
“I think that's so cool. I'm such a huge fan of like the cozy software movement where you like make software tools for like five of your friends and you guys use it together.”
Listen at 1:13:53
Users should try ChatGPT Work on the web and desktop app.
“They should use ChatGPT in the web and try work.”
Listen at 1:19:22
ChatGPT Work can continue executing tasks in the cloud while users are offline.
“You've finally got these things running in the cloud doing real work.”
Listen at 1:20:28
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Featuring
ChatGPT
Barbarian Days
The Power BrokerBooks & mentions
Listen to the full episode and explore every guest, topic, and moment on PodLume.