You’ve used AI to answer questions, write emails, and maybe draft a blog post. That’s all on demand: you ask, it answers, you move on. But there’s a new kind of AI that doesn’t wait for you. It takes a project on Monday and works on it all week, checking in when it hits a decision point. Those are persistent AI coworkers, and OpenAI’s product lead just called them the third era of AI.
Here’s what that shift means, why it matters for your work, and how to try one this week. No hype, just the practical version.
AI’s first two eras (and why the third is different)
Think of AI in three phases, because that’s exactly how OpenAI’s Tara Seshan frames it. She runs product for Codex and ChatGPT Work, two of the biggest agent products on the market, so her framing is worth stealing.
Era one was the chatbot. You typed a question, got an answer, end of story. ChatGPT in its early days, basically. Useful, but it was a tool you had to drive.
Era two brought agents on demand. You gave a task, the AI planned it out and executed it while you watched, then stopped. Think of asking an agent to write a report and coming back twenty minutes later to read it. You were still the supervisor in the loop for everything.
Era three is the persistent AI coworker. The AI gets a goal and works on it continuously, across days, integrating with your tools and your calendar. It sends you updates, asks for your judgment at the right moments, and keeps going while you sleep. That’s a fundamentally different relationship, and it changes what you actually do all day.
What a persistent AI coworker actually does
Concrete example, because this is easier to show than explain. Say you’re planning a product launch in a month.
In era two, you’d ask an agent to research competitors, get the report, then ask another agent to draft the landing page copy, get that, then ask a third to schedule the posts. You’re the project manager, chaining tasks together by hand.
With persistent AI coworkers, you’d hand the whole launch to a small team of agents with one standing goal. One tracks competitor pricing and flags changes. Another drafts campaign copy and revises it against your feedback. A third monitors the launch metrics as they come in and sends you a daily digest. You review, steer, and approve. They do the grinding.
Ambitious people are already running this playbook. A solo founder we covered recently runs 15 AI agents at once to handle engineering, customer support, and investor updates, and he’s one person with a handwritten list.
“Steering vs rowing”: your new job description
Here’s the mental model that makes all of this click, and it comes straight from the Lenny’s interview: the era of rowing is over, and the era of steering has begun.
Rowing means doing the execution: writing the first draft, pulling the data, sending the follow-ups. Steering means deciding where the boat goes: picking the goal, setting the quality bar, making judgment calls, and owning the outcome. AI is getting brutally good at rowing. So the human job is becoming pure steering.
And that’s a scary sentence if your value right now comes from being a great rower. But here’s the flip side, which is also from the interview: if execution is cheap, then ambition becomes the bottleneck. The person who picks bigger, bolder goals and steers them well suddenly has an edge that didn’t exist before. Judgment and taste are the new scarce resources, and those are exactly the skills you already have as a human.
Real products you can try today
You don’t need to wait for the future, because three products already work this way, and two of them are basically free to start.
- Codex (chatgpt.com/codex) is OpenAI’s persistent coding agent. It runs background tasks, fixes its own errors, and reports back. We wrote a full beginner guide on Codex persistent agents, including how to set standing tasks that run while you sleep.
- ChatGPT Work (openai.com/chatgpt-work) is the non-technical version. It handles documents, research, and ongoing projects with the same persistent model, and it’s aimed squarely at people who never want to see a line of code.
- Claude Cowork is Anthropic’s answer, an AI workspace where multiple agents collaborate on your behalf. We compared it against ChatGPT Work head to head, and the honest answer is that both are good at different things.
Start with one standing task, like a weekly competitor digest or a research brief. Set it up once, check the first result, correct it, and let it run. That’s the entire onboarding process.
What this means if you’re not technical
This is the part that should reassure you: persistent AI coworkers are not a coding thing. The whole point of ChatGPT Work and Claude Cowork is that you talk to them like a colleague, not like a computer.
You’ll still want basic skills, though, and they’re the same skills a good manager has: writing clear goals, giving specific feedback, and checking work instead of assuming it’s right. If you can brief a junior colleague, you can brief an AI coworker. In fact, that’s the exact skill transfer the interview highlights.
For small business owners, this is the biggest opportunity. A dog boarding owner we profiled runs her whole operations with Claude because a persistent agent handles the scheduling and follow-ups that used to eat her evenings. That’s the pattern: the AI does the rowing, she does the steering, and the business runs without her burning out.
How to brief your first AI coworker
Ready to try this without getting in over your head? The briefing process is simpler than you think, and it’s the same skill set you’d use with a new hire.
- Pick one boring, repetitive task that takes you an hour or more each week. A weekly report, a research digest, a recurring spreadsheet. That’s your pilot project.
- Write the brief like an email to a smart assistant. Include the goal, the inputs (which files or links matter), the output format you want, and one example of what “good” looks like.
- Run it once while you watch. Check the first output line by line. Correct mistakes in plain language: “the numbers should come from the Q2 file, not the Q3 one.”
- Then let it run on its own and review the results daily for the first week. Once the output is consistently usable, you have your first persistent AI coworker.
The founders and operators doing this best all say the same thing: the first brief is the hard part, and everything after that is just steering.
The honest downsides
Okay, the counterweight, because every new thing has one. Persistent AI coworkers fail loudly, embarrassingly, and sometimes expensively.
They make confident mistakes, they fabricate details, and they can spin their wheels on a bad approach for hours before you notice. That’s why the steering job is not optional. You still need to review outputs, spot-check sources, and kill bad directions early. The four roles framework we published covers exactly where AI belongs in a workflow and where it should sit out.
Cost is the other real issue. Persistent agents consume tokens around the clock, and token bills can surprise you if you don’t set limits. Start small, one agent, one task, and scale only when you trust the output.
Takeaway
Persistent AI coworkers are the closest thing to a new hire you can get for a few dollars a day. They’re not magic, they’re not perfect, but they change the math on what one person can do in a week.
Try the steering mindset with one standing task this week. Brief it like an employee, review its work like a manager, and see if your calendar gets some breathing room. Mine did.
