AI Agents Are Becoming Coworkers Who Stay on the Job
OpenAI product lead Tara Seshan says the next step is people and agents doing work together, not one by one.
Lenny's Podcast: Product | Growth | Career · AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)Tara Seshan, a product lead at OpenAI, was asked how people inside the company are working with AI in ways the rest of us might soon copy. She answered with a short history and a prediction.
"If you think about the first era of AI products as chat, the second era of these products working with agents," Seshan said. "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?"
Seshan means something more lasting than a chatbot that answers and forgets, and more broad than the coding helpers that define much agent use today. This coworker would keep context, do long stretches of work on its own, and check back for direction.
The clearest sign came from inside OpenAI itself. People started posting screenshots of their Codex threads in Slack to show colleagues how they reached a result. A thread is the written record of back and forth with the agent.
That habit showed the limit, Seshan argues. Each person was working one on one with an agent, apart from what colleagues were doing with theirs. Passing around pictures of that work was useful but clumsy, not a natural way to work together.
Her answer is to give agents bigger jobs with less step by step instruction, then step in to steer them after they have worked alone for a while. Seshan describes the ideal as "work feels like a multiplayer game," with people guiding their agents while the agents take care of the routine, step by step work. Lenny Rachitsky, who interviewed her, pictured it this way: Seshan's agent does some analysis, then Rachitsky's agent checks it.
The challenge, she adds, is bringing that power to the billion people who use ChatGPT mainly for chat, without making them think about harnesses. That is the insider term for the behind the scenes setup that lets agents act in other software.
Questions people ask
Would a persistent coworker keep working when I am offline?
Yes, if it is connected to cloud software and has permission to act. The idea is you assign work, go offline, and return to results and questions, rather than watching each step.
Do I still need to check what the agent did?
Yes. Because knowledge work cannot be checked from the final answer alone, the agent needs to show sources, steps, and reasoning. You still review, redirect, and take responsibility for the outcome.
Will using these agents require learning to code?
No. The goal Seshan describes is that ordinary ChatGPT users get agent help without choosing modes or learning new terms. The complexity stays hidden, while the user gives goals and feedback.