I really enjoyed the essay. One of the places I wanted to start is the core idea of what agency is. The term "agent" has become the de facto moniker for what we're doing with AI right now, agentic AI. You defined agency as this raw determination to make things happen without waiting for permission. When you think about why AI doesn't fit that bill, what's your thought process? What is lacking, such that "agent" is not really the right term?

I've been thinking about this for a while now, because, as you said, "agent" as a word has been thrown around a lot recently. I work for a company whose main feature is Replit Agent, right? But I've grown to dislike the term, because fundamentally the way I characterize agency is as a psychological trait. It's this idea of disregarding permission and not waiting for validation before making things happen.

And it's different with AI, to me at least, because of what I've seen happen when you try to make software agents, artificial agents, more like that: more proactive, more independent. People don't like that at all. My personal definition is that they are not doing what they're told, by definition. That's exactly what they do.

100%. If you look at many of the industries where AI has had the biggest impact, one that I've spent a decent amount of time digging into is defense, where, to your point, people don't want fully agentic AI making kill or no-kill decisions. People want human-in-the-loop decisions. The point is to augment them to do what they do best. As AI progress continues, do you think we get to the point where people want to unleash fully agentic AI on the world? Or do we stay in this paradigm where people always want some human agent in the loop, making decisions and pointing the technology in a specific direction?

To me, it boils down to two variables. The first is how comfortable we are with risk, and the second is how good the technology is. At one end of the spectrum, you have basically infallible technology. On the risk side, you have basically zero tolerance for risk. The typical example is defense, or even self-driving cars. To me, that's a very good example of technology that is borderline perfect right now, at least in San Francisco. It feels like it.

So to me it boils down to this: is the variance of the model's success too large for our risk tolerance? If that's the case, we will want to see human accountability. I go back all the time to the idea of plane autopilot. We still want to have at least two pilots on deck all the time, even though we do have autopilot for these planes, because we don't want planes to go down.

I wanted to dive a little deeper into people with high agency. Obviously, their end goal is to pursue something without waiting for permission. But what are some of the common characteristics you've seen in the people who are successful with AI?

One thing I haven't been able to fully expand on in the essay is the notion of what I think the industry right now calls taste. It's basically the ability to judge a good output from a bad output. The reason I haven't fully expanded on it is that, to me, it's downstream of agency. I still think that the ability to produce a bunch of results, let's say you want to create a website, the sheer determination of making a hundred iterations, a thousand iterations, for days and days and days, is to me at least a precondition to then being able to pick the best one.

But then again, these people also tend to have some level of understanding of what could work. There's this viral clip of Rick Rubin where he says, "I have no talent whatsoever. At least I know what is good. And I'm decisive about it." To me, that's all there is to say about high-agency people. They know they can do it. They have no CS degree. They have no preparation. They know zero about coding or design or marketing, and still, deep down, they're like, "Yeah, a website, I can do it. I can make a web app. I can be a founder, even in tech." And then they have the decisiveness of picking what they believe the market wants, and almost the irrationality of being out there and trying to sell it. And usually they find out that they are actually successful.

I love all of these examples. Recruiting is a great example. There are a bunch of people who personify that, who have honed their craft over the course of many years. You made the point that when you were building your first company, it took you nine months to get from an idea to an early prototype. That was your iteration cycle. Now those cycles can happen much faster. With people like MrBeast, it was 10 years of publishing a YouTube video every 10 days. That's how you hone taste. The difference is that now, instead of spending 10 years going through that number of iterations, you can go through them much more quickly. So it makes perfect sense that taste comes as agency is developed. You also talked in the piece about how high-agency people are often attracted to a specific mechanism, which is startups, where there's a very low floor and a high ceiling.

And low structure, because to me the interesting thing about these people is that they don't like to be managed. I don't want to make the point that they can only be founders, albeit typically they are. It's the typical 10x engineer, right? But the shape they have today, at least within tech, is maybe an engineer that spans from data to front end, or a marketer that is totally willing to start tweaking prompts in the AI infrastructure of the company they work for. It's almost an expansion of the concept of range. But to me, again, it's a consequence. The idea is: I have a goal in mind, and I'm willing to do whatever it takes to make it happen.

So typically that's the kind of impact you can have right now, mostly in tech. There's less red tape, an extremely high ceiling, as you said, and low structure that gives you latitude you otherwise wouldn't get in a typical corporate, old-style industry environment.

It feels like ownership is a big piece of it. People want to feel ownership of what they're building, even if they're not the founder of the company. They know they have ownership over the output. You made this point in the essay as well: it feels like the fulfillment of Sam Altman's prophecy in 2023 that we'll start to see one-person billion-dollar companies. We are seeing a lot of companies scale to $100 million-plus of revenue with four or five people, much more efficiently than before. Do you feel like that is a scalable motion? You mentioned a lot of examples in the essay of people who are very much not professional software developers being able to create software. Does that become the status quo, where almost everybody who is high agency is able to customize software to their needs? Or is it a very select group of people capable of building very specific types of tools that can scale like that?

What I'm betting on is the former: that we're seeing the democratization of making software, to an extent. It's also not going to be a clean transition. It's definitely going to be messy. We're going to see, over and over, companies built on brittle ground because the founder was not technical, was rushing to market, and was building without the necessary foundations. We're going to see a bunch of stories like that, maybe companies getting hacked, or companies not being able to withstand the traffic and the scale. So I'm not saying this is going to be clean.

But at the same time, to me it feels like the early internet. The general emotion of the market is such that companies are more efficient. That's obvious, at least from my vantage point, working for a company that is currently serving these customers. I'm seeing absolutely insane things. The barbershop example to me is incredible. There's this barbershop that is creating a waiting list and appointment tracking, all on Replit, and these people have absolutely no knowledge of code. So it doesn't necessarily need to be Midjourney, half-a-billion-a-year kind of scale. It can be much smaller than that: just a small or medium enterprise entrepreneur who wants to make their business more efficient.

To me, that's definitely a shift in how software operates, and therefore in the entire economy, because I still believe that software did eat the world. So it only follows that if the cost of making software is going down like that, it must necessarily have some kind of impact on the economy.

It feels like the end result of something that's been happening for a long time, which is the death of credentialism. Before, it was, "I'm allowed to because I have a degree, because I've gone to the school, I've gotten the training." Now it's just a function of whether you're willing to do it, regardless of your credential. What do you think set off that move away from credentialism? Because it happened even before the current wave of AI. What has changed about people's mentality?

It is clear that there's a larger move toward decentralization. That's happening in media, for instance. You used to go to The New York Times and CNN to get your news, and now you have individual media, Joe Rogan or Kyle right now, that are democratizing access to media. It feels like, for the first time, we can have someone with the level of experience of a PhD do the job for us. And that is necessarily going to move us further away from credentialism.

Because fundamentally, a point I'm trying to make in the essay is that credentialism is a form of local monopoly. And the reason it's a monopoly is that it takes years to get there. If that's no longer needed, because you can literally have software do that for you, well, that's going to change everything, necessarily. And that, to me, is the end of the wave of the internet that started in the '90s.