What AI does is it makes skilled human expertise available to everybody for cheap. And you would think that that would replace experts. And the reality is it actually increases the demand for experts. What I want to do for the rest of the video is unpack this paradox. What actually happens after automation? And why is it that it seems like for companies like ours, which are supposedly at the edge of automating everything, there's actually a ton of human work to do or hiring tons and tons of humans, even as we use a lot of agents. What actually is going on there? Because look at Dario. Dario's out there being like, AI could wipe out over 50% of all entry-level white-collar jobs. Ken Griffin, he runs Citadel, which is a gigantic hedge fund. He's sitting up on stage talking about seeing an agent do some work. And he says, these are not mid-tier white-collar jobs. These are extraordinarily high-skilled jobs being automated by a genetic AI. If even people like him are experiencing these frontier models and going, holy shit, I don't know if there's going to be work in the future. And then if you look at the public polling, a lot of people are very scared of it. And I think that's super at odds with my actual experience using this stuff. So that's what I want to do with this video. And there's a companion explainer. If you want to read something really in-depth and read it with your agent, you can go there. We'll put a link in the description for you. So what I want to do first is I just want to talk a little bit about how do we actually use agents? How have agents landed in our day-to-day work inside of a company of 25 people? I think of us as being this microcosm of what is possible if you're trying to push this to the maximum extent possible. And then we'll talk about, based on this, why are things the way that they are? And what does it mean for the future of work for knowledge workers, especially highly skilled knowledge workers? Just a little background on me and Every. I run this company called Every. We've been around for about six and a half years. We've been covering AI and agents and how agents might change work since 2022. So for a long time, since before ChatGPT came out. And three years ago, I wrote this piece about what I called the allocation economy. And in that piece, and this is back when agents weren't really a thing, it was just ChatGPT doing little responses or maybe co-pilot. In that piece, what I argued is that the way that we will probably work with agents in the future is we will work with them like managers do. So the best corollary is going to be human manager skills. And if you're a good human manager, you learn to do things like delegate or be able to know when to micromanage or be able to split a task up into discrete units that you can give to different people to get a task done. Those are the kinds of skills. And it's not an exact transfer, but that's the best analogy for the kind of skill that's going to be most valuable in AI. I wrote that in 2023 or so. We'll link to this as well. And then about a year ago, I got super Claude Pilled. And Kieran Claussen, who also works at Every, who's the GM of Quora, was suddenly coding with Claude Code. This is in the maybe Sonnet 3.7 era was when we were first, holy shit, this thing is really working. He was suddenly coding without looking at the code. He was just in a CLI all day. I started using it for my own coding. And then I started using it for just regular knowledge work. And I was like, holy shit, I think this is the template for how you're going to do knowledge work going forward. I actually went on the Lenny podcast about 12 months ago, and I predicted that Claude Code is the most underrated tool for knowledge work. And I feel like a lot of that has started to happen. So because of the kind of company that we are, everybody internally is really AI pilled. We're all early adopters. And we all get to test models before they come out, because we're good at testing them. We're good at telling the model companies, this is what it's good for. This is what it's not good for, that kind of stuff. I think we get to be a little bit of a lab for what is work going to look like in the same way that we've been able to see maybe how working with agents might feel three years ago. We're getting a lens into that now that I want to share. I think there's major implications for this paradox we're talking about. So when I step back and think about how is work with agents starting to happen? I think there's two main shapes. And there's one shape that was very well predicted by the discourse and one shape that was not. The one that is very well predicted by the discourse is it is very common right now for us to, when we have work we want to do, just at mention someone in Slack and say, hey, can you do this work for me? And that's not a person. It's a bot. It's an AI. We use a bunch of different AIs. Some of them are built internally. So we have a couple. We have one called Claudie, which is built internally. We have a couple of videos about that. Claudie runs our whole consulting business. We've got one called Andy, which is an editorial assistant. We also have one called Victor that we've been using. Victor just raised a massive round and they're doing a really good job with this kind of thing. So we have that form factor. Delegate to an agent over Slack. It just goes and does things. It does everything from marketing and brand research to recording the A-B tests for our YouTube thumbnails to sending out our consulting proposals to clients to putting together the first draft of decks for our clients. There's tons of stuff that we do with that. So that's one service. The other surface, which I think is weirder and more important, is there is starting to become with Codex and Cowork, but I think Codex is the furthest along on this paradigm. We're starting to see that what I would call agent orchestration software, which is basically what Codex is or what Cloud Code is or whatever, where each chat spins off an agent that's on your computer. That's becoming the operating system for work, where all the work that I do is in Codex or Cowork. And it's an agent that lives on my computer that I access through a GUI that has access to all of my computer, that is connected to all my stuff, that, and this is really important, has a browser in it. I do everything from my inbox. I write in there, do a lot of coding in there. When the quarter ends for the business, I analyze our P&L and that kind of stuff. So these are the two shapes of work. One is delegation and one is this agent orchestration system where you're sitting in there and you have multiple agents going at once. You're watching them work. You're very in the loop with them and they're very in the loop with you. It's a true AI human collaboration type system. And I think those are both really interesting because what we find for both of these, the only way that it works is if a human is involved. One example, we have Claudie. Claudie runs our consulting business, like I said. Claudie works because there's a guy named Nitesh, who's one of our senior AI engineers. And Nitesh's job is basically just to see all the places in which Claudie isn't quite working and then just be constantly helping Claudie work better. And it's not one of those things where he's the puppet master and he's actually doing all the work. Claudie is actually giving him and the business a ton of leverage. But I think there's this misconception that once you set up an agent, it just works. And the reality is very different. The further away it's been from a human helping to correct it and helping to make sure that it's doing the right thing, the worse it gets. And you can see the same thing in this agent orchestration system where when I have to do real work, it's so much better when I'm in the loop with my agent, when I'm on my computer in Codex. And that human collaboration bit is the thing that makes it so powerful. So why do we need more humans? If that's the case, let's get into that. So hopefully now what we've done is we have identified this paradox. We've gone into here's how work works with AI. You're either delegating fully to something in Slack or you're going back and forth with a bunch of different agents on your computer. But both of those require you or someone else in some way to be intervening a lot. Okay, so why does this then make more work for humans? Let's go through the argument. One, AI makes yesterday's human competence cheap. Language models are trained on data. Data is the visible residue of competence. An example, the ability to write a pull request to submit code to a code base used to be rare and expensive. Now, literally anyone can do it. If you have not done it, you should just go do it right now on Cloud Code or Codex or whatever. It's now extremely easy. When you make competence cheap, when you make yesterday's competence cheap, adoption skyrockets. Everyone wants to do it. Everyone's making pull requests. Everyone's making images. Everyone's tweeting. The demand for cheap competence is super high. And so it pervades everywhere. And we see this internally all the time. Suddenly you have ops people and customer service people who are submitting pull requests or you've got marketers doing YouTube thumbnails, but then the designers are like, whoa. And what happens when you have all those people who want to do something and suddenly can do it with this cheap competence, you get this glut of new work. It's generic. It ranges from anywhere from it's complete slop to it's a decent start. We can work with this. But if you're not actually good at what you do and you're using these tools to just do the default thing, you're going to make something that's not that great. It's going to be more than you could do before. And that's amazing. But using it by itself with a default, the laziest prompt possible is not going to actually produce work that's valuable. There's a very popular term now: slop. And the thing that I think is important to understand is slop is not any one particular thing. It's not, oh, it just uses em dashes too much. We can get rid of that. It's not a certain sentence rhythm. It's not that Claude just uses purple accents too much on landing pages. It's what humans start to notice when you now have a tool that by default can produce the same kind of thing in a lot of different circumstances. We start to notice that and we start to smell a rat. Everything looks a little bit similar. When everyone has access to this tool with cheap competence and you have this glut of new things on the market that are all sort of the same, you actually increase the demand for difference. And who do they call? Experts. What experts can do, for example, when anyone can write a pull request, is they can come in and say, yeah, I see the problem that you're trying to solve and this is a good first draft, but then they can take it and turn it into something that can actually be merged and deployed to production. That's why a lot of experts are, I don't like AI because it puts way more work on my plate to clean up all this slop. But there's a couple of different responses to what do experts do now in this world? One is they get annoyed. But two, and it was really interesting is now a lot of experts create systems to be able to absorb and leverage this new work. So internally, for example, at Every, I can write pull requests for our products. The default pull request I'm going to write is probably not that useful. But now we have systems. So that can be your cloud MD file or your repo rules, or just a social contract for how and when I submit pull requests. For example, one thing I do is I just make my own little vibe coded version so I can show people and then it gets integrated into the tool later and into our product later. So one thing experts do is they build systems and processes to handle all this new work and turn it into something amazing. Or what they can do is take this yesterday's competence that is now available to everyone and use it themselves and stand on that as the floor and use it to do things that they would never have been able to do. For example, internally for us, we have several people who each runs an entire software product by themselves. That completely would not be possible before. There's lots of examples of that. And so that is why in practice, in skilled knowledge work domains, it actually means that there's a lot more work to do because every agent requires a human to be good. Everyone having access to yesterday's competence increases the amount of stuff that gets produced. You need experts to come in to help you make the stuff that gets produced and turn it into something that's not sloppy. One way to think about this is there's almost a Zeno's paradox going on here where in Zeno's paradox, you've got Achilles. Achilles is racing a tortoise. The tortoise is ahead of Achilles, starts ahead because the tortoise is super slow. And the paradox asks, how could Achilles ever catch the tortoise? Every time that he goes halfway to the tortoise, the tortoise moves a little bit and he can never close the distance. And there's something like that going on with AI where we're the turtle and we start 50 yards ahead and AI is hoovering everything up and it's starting to nip at our heels. And the question is, is it going to catch us? I really don't think this is going to happen. I have a much longer argument about how this all works. So if you care about a lot of the details, I would go read the article. But let's assume that we just have this super smart AI. This is my definition of AI. It's something that you never turn off. It's always working. It's always running. Once you do that and it's economically valuable for you and you pay to keep it on all the time, I think that's AGI. It's open ended. So it's finding new goals and doing new stuff. This is why in practice, when you actually look at this, it does not eliminate human work. None of that work is really valuable unless a human being is involved. And this is not to say that jobs aren't going to change. My job is completely different and there's lots of different parts of the economy. So the stuff that we see internally at Every, it may apply in different ways to different kinds of jobs. But what I wanted to do here is take the big fear that I think a lot of people have and subject it to the test of what actually do we see and what have we seen over the last couple of years. And companies are really trying to do this well. And I think we see a clear story here that I think if people understood that better, it would change the discourse. And you may be sitting here being like, cool, I guess I understand that now, but what do I do? And my big thing, I think the biggest thing really is just learn how to use them. Just ride the models. As the models get better, you get more powers. You get more cheap human competence packaged up that you can use for whatever you want. If you're a smart, curious person who wants to do interesting work, it's a super tool for you to go do all that stuff. There's going to be a lot of demand for you to do more of it. If you're interested in this, we have a ton more on Every about this. I've got a whole section about the benchmarks. What does it mean when the benchmarks go up? What happens when we hit AGI? So if you have questions about all that stuff, we have a lot on Every for you to read. And I'd love to hear from you. This is a complicated topic. So please subscribe, please comment. Tell me if you disagree. Let's have a discussion about it. And remember, just ride the models.
think that that would replace experts. And the reality is it actually increases the demand for experts. What I want to do for the rest of the video is unpack this paradox. What actually happens after automation? And why is it that it seems like for companies like ours, which are supposedly at the edge of automating everything, there's actually a ton of human work to do or hiring tons and tons of humans, even as we use a lot of agents. What actually is going on there? Because look at Dario. Dario's out there being like, AI could wipe out over 50% of all entry-level white-collar jobs. Ken Griffin, he runs Citadel, which is a gigantic hedge fund. He's sitting up
on stage talking about seeing an agent do some work. And he says, these are not mid-tier white-collar jobs. These are like extraordinarily high-skilled jobs being, I'm going to pick a word, being automated by a genetic AI. If even people like him are experiencing these frontier models and going like, holy shit, I don't know if there's going to be work in the future. And then if you look at the public polling, a lot of people are very scared of it. And I think that's super at odds with my actual experience using this stuff. So that's what I want to do with this video. And there's a companion
explainer. Like if you want to read something really in-depth and read it with your agent, you can go there. We'll put a link in the description for you. So what I want to do first is I just want to talk a little bit about how do we actually use agents? How have agents landed in our day-to-day work inside of a company of 25 people? I think of us as being this microcosm of what is possible if you're trying to push this to the maximum extent possible. And then we'll talk about, okay, based on this, why are things the way that they are? And what does it mean for the future of work for knowledge
workers, especially highly skilled knowledge workers? Just a little background on me and Every. I run this company called Every. We've been around for about six and a half years. We've been covering AI and agents and how agents might change work since 2022. So for a long time, since before ChatGPT came out. And three years ago, I wrote this piece about what I called the allocation economy. And in that piece, and this is back when agents weren't really a thing, it was just ChatGPT doing little responses or maybe co-pilot. In that piece, what I argued is that the way that we will probably work with agents in
the future is we will work with them like managers do. So the best corollary is going to be human manager skills. And if you're a good human manager, you learn to do things like delegate or be able to know when to micromanage or be able to split a task up into discrete units that you can give to different people to get a task done. Those are the kinds of skills. And it's not an exact transfer, but that's the best analogy for the kind of skill that's going to be most valuable in AI. I wrote that in 2023 or so. We'll link to this as well. And then about a year ago, I got super fucking
Claude Pilled. And Kieran Claussen, who also works at Every, who's the GM of Quora, was suddenly coding with Claude Code. This is in the maybe Sonnet 3.7 era was when we were first like, holy shit, this thing is really working. He was suddenly coding without looking at the code. He was just in a CLI all day. I started using it for my own coding. And then I started using it for just regular knowledge work. And I was like, holy shit, I think this is the template for how you're going to do knowledge work going forward. I actually went on the Lenny podcast about 12 months ago,
and I predicted that Claude Code is the most underrated tool for knowledge work. And I feel like a lot of that has started to happen. So because of the kind of company that we are, everybody internally is really AI Pilled. We're all early adopters. And we all get to test models before they come out, because we're good at testing them. We're good at telling the model companies, this is what it's good for. This is what it's not good for, that kind of stuff. I think we get to be a little bit of a lab for what is work going to look like in the same way that we've been able to see maybe how working
with agents might feel three years ago. We're getting a lens into that now that I want to share. I think there's major implications for this paradox we're talking about. So when I step back and think about how is work with agents starting to happen? I think there's two main shapes. And there's one shape that was very well predicted by the discourse and one shape that was not. The one that is very well predicted by the discourse is it is very common right now for us to, when we have work we want to do, just at mention someone in Slack and say, hey, can you do this work for me? And that's not a person.
It's a bot. It's an AI. We use a bunch of different AIs. Some of them are built internally. So we have a couple. We have one called Claudie, which is built internally. We have a couple of videos about that. Claudie runs our whole consulting business. We've got one called Andy, which is an editorial assistant. We also have one called Victor that we've been using. Victor just raised like a massive round and they're doing a really good job with this kind of thing. So we have that form factor. Delegate to an agent over Slack. It just goes and does things. It does everything from marketing and brand research
to recording the A-B tests for our YouTube thumbnails to sending out our consulting proposals to clients to putting together the first draft of decks for our clients. There's tons and tons of stuff that we do with that. So that's one service. The other surface, which I think is weirder and more important, is there is starting to become with Codex and Cowork, but I think Codex is the furthest along on this paradigm. We're starting to see that what I would call agent orchestration software, which is basically what Codex is or what Cloud Code is or whatever, where each chat spins off an agent that's on
your computer. That's becoming the operating system for work, where all the work that I do is pretty much in Codex or Cowork. And it's an agent that lives on my computer that I access through a GUI that has access to all of my computer, that is connected to all my stuff, that, and this is really important, has a browser in it. I do everything from my inbox. I write in there, do a lot of coding in there. When the quarter ends for the business, I analyze our P&L and they're like, all that kind of stuff. So these are the two shapes of work. One is delegation and one is this agent orchestration
system where you're sitting in there and you have multiple agents going at once. You're watching them work. You're very in the loop with them and they're very in the loop with you. It's like a true AI human collaboration type system. And I think those are both really interesting because what we find for both of these, the only way that it works is if a human is involved. One example, we have Claudie. Claudie runs our consulting business, like I said. Claudie works because there's a guy named Nitesh, who's one of our senior AI engineers. And Nitesh's job is basically just to see all the places in which Claudie
isn't quite working and then just be constantly helping Claudie work better. And it's not one of those things where he's the puppet master and he's actually doing all the work. Claudie is actually giving him and the business a ton of leverage. But I think there's this misconception that once you set up an agent, it just works. And the reality is very different. The further away it's been from a human helping to correct it and helping to make sure that it's doing the right thing, the worse it gets. And you can see the same thing in this agent orchestration system where when I have to do real
work, it's so much better when I'm in the loop with my agent, when I'm on my computer in codex. And that human collaboration bit is the thing that makes it so powerful. So why do we need more humans? If that's the case, let's get into that. So hopefully now what we've done is we have identified this paradox. We've gone into here's how work works with AI. You're either delegating fully to something in Slack or you're going back and forth with a bunch of different agents on your computer. But both of those require you or someone else in some way to be intervening a lot. Okay, so why does this then make
more work for humans? Let's go through the argument. One, AI makes yesterday's human competence cheap. So language models are trained on data. Data is the visible residue of competence. And so an example, the ability to write a pull request to submit code to a code base used to be rare and expensive. Now, literally anyone can do it. If you have not done it, you should just go do it right now on cloud code or codex or whatever. It's now extremely easy. When you make competence cheap, when you make yesterday's competence cheap, adoption skyrockets. Everyone wants to do it. Everyone's making pull requests. Everyone's making images. Everyone's tweeting a bunch of stuff.
The demand for cheap competence is super high. And so it pervades everywhere. And we see this internally all the time. Suddenly you have like ops people and customer service people who are submitting pull requests or you've got marketers doing YouTube thumbnails, but then the designers are like, whoa. And what happens when you have all those people who want to do something and suddenly can do it with this cheap competence, you get this glut of new work. That's kind of generic. It ranges from anywhere from its complete slop to it's a decent start. We can work with this. But if you're not actually good at what you do and you're using these tools to just do the default
thing, you're going to make something that's not that great. It's going to be more than you could do before. And that's amazing. But using it by itself with a default, the laziest prompt possible is not going to actually produce work that's valuable. There's a very popular term now slop. And the thing that I think is important to understand is slop is not any one particular thing. It's not, oh, it just uses em dashes too much. We can get rid of that. It's not a certain sentence rhythm. It's not that Claude just uses purple accents too much on landing pages. It's what humans start to notice when you now have a tool that by default can produce the same kind of
thing in a lot of different circumstances. We start to notice that and we start to kind of smell a rat. Everything looks a little bit similar. When everyone has access to this tool with cheap competence and you have this glut of new things on the market that are all sort of the same, you actually increase the demand for difference. And who do they call? Experts. What experts can do, for example, when anyone can write a pull request, is they can come in and say, yeah, I see the problem that you're trying to solve and this is maybe a good first draft, but then they can take it and turn it into something that can actually be merged and deployed to production.
That's why a lot of experts are like, I don't like AI because it puts way more work on my plate to clean up all this slop. But there's a couple of different then responses to what do experts do now in this world? One is they get annoyed. But two, and it was really interesting is now a lot of experts create systems to be able to absorb and leverage this new work. So internally, for example, at every, I can write pull requests for our products. The default pull request I'm going to write is probably not that useful. But now we have systems. So that can be your cloud MD file or your
repo rules, or just a social contract for how and when I submit pull requests. For example, one thing I do is I just make my own little vibe coded version so I can show people and then it gets integrated into the tool later and into our product later. So one thing experts do is they build systems and processes to handle all this new work and turn it into something amazing. Or what they can do is take this yesterday's competence that is now available to everyone and use it themselves and stand on that as the floor and use it to do things that they would never have been able to do. For example,
internally for us, we have several people who each of them runs an entire software product by themselves. That completely would not be possible before. There's lots of examples of that. And so that is why in practice, in skilled knowledge work domains, it actually means that there's a lot more work to do because every agent requires a human to be good. Everyone having access to yesterday's competence increases the amount of stuff that gets produced. You need experts to come in to help you make the stuff that gets produced and turn it into something that's not sloppy. One way to think about
this is there's almost like this Xeno's paradox going on here where in Xeno's paradox, you've got Achilles. Achilles is racing a tortoise. The tortoise is ahead of Achilles, starts ahead because the tortoise is super slow. And the paradox asks, how could Achilles ever catch the tortoise? Every time that he goes halfway to the tortoise, the tortoise moves a little bit and he can never close the distance. And there's something like that going on. Feels like an AI where we're like the turtle and we start 50 yards ahead and AI is just like hoovering everything up and it's starting to nip at our
heels. And the question is, is it going to catch us? I really don't think this is going to happen. I have a much longer argument about how this all works. So if you care about a lot of the details, I would go read the article. But let's assume that we just have this super smart AI. This is my definition of AI. It's something that you never turn off. It's always working. It's always running. Once you do that and it's economically valuable for you and you pay to keep it on all the time, I think that's AGI. It's open ended. So it's finding new goals and doing new stuff. This is
why in practice, when you actually look at this, it does not eliminate human work. None of that work is really valuable unless a human being is involved. And this is not to say that jobs aren't going to change. My job is completely different and there's lots of different parts of the economy. So the stuff that we see internally at every, it may apply in different ways to different kinds of jobs. But what I wanted to do here is take the big fear that I think a lot of people have and subject it to the test of what actually do we see and what have we seen over the last couple of years. And companies
are really trying to do this well. And I think we see a clear story here that I think if people understood that better, it would change the discourse. And you may be sitting here being like, cool, I guess I understand that now, but what do I do? And my big thing, I think the biggest thing really is just learn how to use them. Just ride the models. As the models get better, you get more powers. You get more cheap human competence packaged up that you can use for whatever you want. If you're a smart, curious person who wants to do interesting work, it's a super tool for you to go do all that stuff. There's going to be a lot of demand for you to do more of it.
If you're interested in this, we have a ton more on every about this. I've got a whole section about the benchmarks. Like what does it mean when the benchmarks goes up? What happens when we hit AGI? So if you have questions about all that stuff, we have a lot on every for you to read. And I'd love to hear from you. This is a complicated topic. So please subscribe, please comment. Tell me if you disagree. Let's have a discussion about it. And remember, just ride the models.