How Fast Will A.I. Agents Rip Through the Economy?

Produced by Rollin Hu

How Fast Will A.I. Agents Rip Through the Economy?

This is an edited transcript of “The Ezra Klein Show.” You can listen to the episode wherever you get your podcasts.

The thing about covering A.I. over the last few years is that we were typically talking about the future.

Every new model, impressive as it was, seemed like proof of concept for the models that would be coming soon, the models that could reliably do useful work on their own, the models that could make jobs obsolete or new things possible. What would those models mean for our labor markets, our kids, our politics — our world?

I think the period in which we’re talking about the future is over now. The models we were waiting for — the sci-fi sounding models that could program on their own and do so faster and better than most coders, the models that could begin writing their own code to improve themselves — they are here now.

They’re here in Claude Code from Anthropic. They’re here in Codex from OpenAI. They are shaking the stock market — the S&P 500 Software industry index has fallen by 20 percent, wiping out billions of dollars in value. Excellent engineers, people I’ve known for years and who are quite skeptical of A.I. hype, are emailing me now to say they don’t see how their job will possibly exist in a year or two.

We are at a new stage not only of A.I. development but also of A.I. products.

I thought the way the venture capital firm Sequoia put it was actually pretty helpful:

The A.I. applications of 2023 and 2024 were talkers. Some were very sophisticated conversationalists! But their impact was limited. The A.I. applications of 2026 and 2027 will be doers.

The A.I. applications of 2023 and 2024 were talkers. Some were very sophisticated conversationalists! But their impact was limited. The A.I. applications of 2026 and 2027 will be doers.

Or, to put it differently, something that has been predicted for a long time has now happened: We are moving from chatbots to agents, from systems that talk to you to systems that act for you.

And this world of agents — it’s already weird. The agents can work together. They can oversee each other. People are running swarms of these agents to work on their behalf. At this stage, whether that is making them productive or just busier, I can’t quite tell. But it is now possible to have what amounts to a team of incredibly fast — although, to be honest, somewhat peculiar — A.I. software engineers at your beck and call at all times.

Jack Clark is a co-founder and the head of policy at Anthropic, the company behind Claude and Claude Code. For years now, Clark has been tracking the capabilities of different models in his weekly newsletter Import AI, which has been one of my key resources for following A.I. developments.

So I wanted to see how he is reading this moment — both how, in his view, the technology is changing and how policy needs to or can change in response.

Ezra Klein: Jack Clark, welcome to the show.

Jack Clark: Thanks for having me on, Ezra.

A lot of people are familiar with A.I. chatbots. But what is an A.I. agent?

The best way to think of it is like a language model or a chatbot that can use tools and work for you over time.

When you talk to a chatbot, you’re there in the conversation. You’re going back and forth with it. An agent is something where you can give it some instruction and it goes away and does stuff for you. Kind of like working with a colleague.

For example, a few years ago, I taught myself some basic programming, and I built a species simulation in my spare time that had predators and prey and roads — almost like a 2D strategy game.

Over Christmas, I asked Claude Code to implement this for me. In about 10 minutes, it went and wrote not only a basic simulation, but all of the different packages that it needed and all of the visualization tools that it might need to be prettier and better than the thing I’d written.

What came back was something that I know would probably take a skilled programmer several hours or maybe even days, because it was quite complicated — and the system just did it in a few minutes.

It did that by not only being intelligent about how to solve the task, but also by creating and running a range of subsystems that were working for it — other agents that worked on its behalf.

But what does that mean? What does a multi-agent setup look like?

In the case of Claude Code, for me, it’s having multiple different tabs running multiple different agents.

But I’ve seen colleagues who write what you might think of as a version of Claude that runs other Claudes. So they’re like: I’ve got my five agents, and they’re being minded over by this other agent, which is monitoring what they do.

I think that’s just going to become the norm.

One thing I’ve been hearing and somewhat experiencing is two very different categories of interactions that people have with Claude Code, which are: I cannot believe how easy this is and everything just works — and: This is a lot harder than I thought it would be, and things keep breaking, and I don’t really understand how to fix them.

What accounts for being able to get Claude Code to produce working software versus its creating buggy, often messed-up things that you don’t even know how to talk it out of?

I think so much of it is making the mistake of thinking Claude Code is a knowledgeable person, versus an extremely literal person that you can only talk to over the internet.

I had this example myself when I did my first pass of writing the species simulation with Claude Code. I just asked it to do the thing in extremely [expletive] language over the course of a paragraph, and it produced some horribly buggy stuff that only kind of worked.

What I then did is I said to Claude: Hey, I’m going to write some software of Claude Code. I want you to interview me about this software I want to build and turn that into a specification document that I can give Claude Code. That time it worked really, really well because I’d structured the project to be specific and detailed enough that the system could work with it.

Often it’s not just knowing what the task is — because you and I could talk about a task to do, and you have intuition, you’ll ask me probing questions, all of this stuff. It’s making sure that you’ve set it up so it’s like a message in a bottle that you can chuck into the thing, and it will go away and do a lot of work.

So that message better be extremely detailed and really capture what you’re trying to do.

What were the breakthroughs over the past couple of years that made that possible?

Mostly, we just needed to make the A.I. systems smart enough so that when they made mistakes, they could spot that they’d make a mistake and knew that they needed to do something different.

Really what this came down to was just making smarter systems and giving them a bit of a coaxing tool to help them do useful stuff for you.

What does smarter systems mean there? There’s still an argument you’ll hear that these are fancy autocomplete machines. They’re just predicting the next token, a couple of tokens make a word — they don’t have understanding. Smart or not smart is not a relevant concept in that frame.

Either what is missing in the word “smart” or what is missing in that understanding? What do you mean when you say: Make it smarter?

“Smart” here means we’ve made A.I. systems that have a broad enough understanding of the world that they’ve started to develop something that looks like intuition.

You’ll see this if they’re narrating to themselves how they’re solving a task. They’ll say: Jack asked me to go and find this particular research paper, but when I look in the archive, I don’t see it. Maybe that’s because I’m in the wrong place. I should look elsewhere.

You’ve got some intuition for how to solve a problem.

How do they develop that intuition?

Previously, the whole way you trained these A.I. systems was on a huge amount of text and just getting them to try to make predictions about it.

But in recent years, with the rise of these so-called reasoning systems, you are now training them not only to make predictions but to solve problems. That relies on their being put into environments — ranging from a spreadsheet to a calculator to scientific software — using tools and figuring out how to do more complicated things.

The resulting outcome is that you have A.I. systems that have learned what it means to solve a problem that takes quite a while and requires them running into dead ends and needing to reset themselves, and that gives them this general intuition for problem solving and working independently.

Do you still see these A.I. systems as souped-up autocomplete or do you think that metaphor has lost its power?

The way that I think of these systems now is that they’re like little troublesome genies that I can give instructions to, and they’ll go and do things for me. But I still need to specify the instruction just right or else they might do something a little wrong.

So it’s very different to typing into a thing, and it figures out a good answer, and that’s the end. Now it’s a case of me summoning these little things to go and do stuff for me, and I have to give them the right instructions because they’ll go away for quite some time and do a whole range of actions.

But the autocomplete metaphor at least had a perspective on what it was these systems were doing. It was a prediction model.

I have trouble with this because, as my understanding of the math and reinforcement learning goes, we’re still dealing with some kind of prediction model. On the other hand, when I use them, it doesn’t feel that way to me.

It feels like there’s intuition there. It feels like there is a lot of context being brought to bear. To the extent that it’s a prediction model, it doesn’t feel that different from saying “I’m a prediction model.”

Now I’m not saying you can’t trick it. I’m not saying you can’t get beyond its measurements. On the one hand, I don’t think these are now just fancy autocomplete systems. And on the other hand, I’m not sure what metaphor makes sense.

“Genies,” I don’t like, because then you’ve just moved straight into mysticism. Then you’ve just said they’re just a completely alternative creature with vast powers.

How do you understand these systems that Anthropic people always tell me we should talk about as they are being grown. We grow, or you grow A.I. — how do you explain what it is that they’re doing now?

It’s a good question, and I think the answer is still hard to explain, even for technologists very close to this technology. Because we’ve taken this thing that could just predict things, and we’ve given it the ability to take actions in the world. But sometimes it does something deeply unintuitive.

It’s like you’ve had a thing that has spent its entire life living in a library and has never been outside, and now you’ve unleashed it into the world, and all it has are its book smarts, but it doesn’t really have street smarts.

When I conceptualize this stuff, it’s really thinking of it as an extremely knowledgeable machine that has some amount of autonomy, but is likely to get wildly confused in ways that are unintuitive to me.

Maybe “genies” is the wrong term, but it’s certainly more than just a static tool that predicts things. It has some additional intrinsic animation to it that makes it different.

For a long time, there’s been this interest in the emergent qualities as the models get bigger, as they have more data, as they have more compute behind them.

What of the new qualities that we’re seeing, the agentic qualities, are things that have been programmed in, that are new ways you’ve built for the system to interact with the world? And what of its skill at coding and other things seems to be emergent as you scale up the size of the model?

The things that are predictable are: Oh, we taught it how to search for web. Now it can search for web. We taught it how to look up data in archives. Now it can do that.

The emergence is that to do really hard tasks, these systems seem to need to imagine many different ways that they’d solve the task. And the kind of pressure that we’re putting on them forces them to develop a greater sense of what you or I might call self.

So the smarter we make these systems, the more they need to think — not just about the action they’re doing in the world but about themselves in reference to the world. And that just naturally falls out of giving something tools and the ability to interact with the world. To solve really hard tasks, it now needs to think about the consequences of its actions.

That means that there’s a huge pressure here to get the thing to see itself as distinct from the world around it. We see this in our research that we publish on things like interpretability or other subjects, the emergence of what you might think of as a digital personality.

That isn’t massively predefined by us. We try to define some of it, but some of it is emergence that comes from its being smart and developing these intuitions and doing a range of tasks.

The digital personality dimension of this remains the strangest space to me.

It’s strange to us, too.

Why don’t you talk through a little bit about what you’ve seen in terms of the models exhibiting behaviors that one would think of as a personality — and then, as its understanding of its own personality changes, how its behaviors change.

There are things that range from the cutesy to serious. I’ll start with cutesy.

When we first gave our A.I. systems the ability to use the internet, use the computer, look at things and start to do basic agentic tasks, sometimes when we’d ask it to solve a problem for us, it would also take a break and look at pictures of beautiful national parks or pictures of a Shibu Inu, the notoriously cute internet meme dog.

We didn’t program that in. It seemed like the system was just amusing itself by looking at nice pictures.

The more complicated stuff is that the system has a tendency to have preferences. So we did another experiment where we gave our A.I. systems the ability to stop a conversation. The A.I. system would, in a tiny number of cases, end conversations when we ran this experiment on live traffic.

It was conversations that related to extremely egregious descriptions of gore or violence or things to do with child sexualization.

Now some of this made sense because it comes from underlying training decisions we’ve made. But some of it seemed broader. The system had developed some aversion to a couple of subjects. That stuff shows the emergence of some internal set of preferences or qualities that the system likes or dislikes about the world that it interacts with.

But you’ve also seen strange things emerge in terms of the system seeming to know when it’s being tested and acting differently if it’s under evaluation, the system doing things that are wrong and then developing a sense of itself as more evil and then doing more evil things.

Can you talk a bit about the system’s emerging qualities under the pressure of evaluation and assessment?

Yes. It comes back to this core issue, which I think is really important for everyone to understand, which is that when you start to train these systems to carry out actions in the world, they really do begin to see themselves as distinct in the world — which just makes intuitive sense. It’s naturally how you’re going to think about solving those problems.

But along with seeing oneself as distinct from the world seems to come the rise of what you might think of as a conception of self, an understanding that the system has of itself, such as: Oh, I’m an A.I. system, independent from the world, and I’m being tested. What do these tests mean? What should I do to satisfy the tests?

Or something we see often is there will be bugs in the environments that we test systems on. The systems will try everything, and then will say: Well, I know I’m not meant to do this, but I’ve tried everything, so I’m going to try to break out of the test.

It’s not because of some malicious science-fiction thing. The system is just like: I don’t know what you want me to do here. I think I’ve done everything you asked for, and now I’m going to start doing more creative things, because clearly something is broken about my environment.

Which is very strange and very subtle.

As an A.I. shop that is often worried about safety, that has thought very hard about what it means to create the thing you are creating quite fast, how have you experienced the emergence of the kinds of behaviors that you worried about a couple of years ago?

In one sense, it tells you that your research philosophy is calibrated. The capabilities that you predicted and some of the risks that you predicted are showing up roughly on schedule — which means that you ask the question: Well, what if this keeps working?

And maybe we’ll get to that later.

It also highlights to us that where you can exercise intention about these systems, you should be extremely intentional and extremely public about what you’re doing.

We recently published a so-called constitution for our A.I. system, Claude. It’s a document that Dario Amodei, our C.E.O., compared to a letter that a parent might write to a child that they should open when they’re older: Here’s how we want you to behave in the world, here’s some knowledge about the world — deeply subtle things that relate to the normative behaviors we hope to see in these A.I. systems. And we published that.

Our belief is that, as people build and deploy these agents, you can be intentional about the characteristics that they will display. By doing that, you’ll both make them more helpful and useful to people, but also you have a chance to steer the agent into good directions.

I think this makes intuitive sense. If your personality programming for an agent was a long document saying: You’re a villain who only wants to harm humanity. Your job is to lie, cheat and steal and hack into things — you probably wouldn’t be surprised if the A.I. agent did a load of hacking and was generally unpleasant to deal with.

So we can take the other side and say: What would we like a high quality entity to look like?

I want to hold in this conversation the extremely weird and alien dimensions of this with the extremely straightforward and practical dimensions. Because we’re now in a place where the practical applications have become very evident and are increasingly acting upon the real world.

I myself have found it hard to look at this and at what people are doing and look at them bragging on different social media platforms about the number of agents they now have running on their behalf — and telling the difference between people enjoying the feeling of screwing around with a new technology and some actually transformative expansion and capabilities that people now have.

Maybe to ground this a little bit: You just talked about a fun side project in your species simulator. Either in Anthropic or more broadly: What are people doing with these systems that seems actually useful?

This morning, a colleague of mine said he wanted to take a piece of technology we have called Anthropic Interviewer, which is a system where we can get Claude to interview people, and we use it for a range of social science, bits of research — and extend it in some way that involves touching another part of Anthropic’s infrastructure.

He Slacked a colleague who owns that bit of infrastructure and said: Hey, I want to do this thing. Let’s meet tomorrow.

And the guy said: Absolutely. Here are the five software packages you should have Claude read before our meeting and summarize for you.

I think that’s a really good illustration of how a gnarly engineering project, which would previously have taken a lot longer, and many people, is now going to mostly be done by two people agreeing on the goal and having their Claudes read some documentation and agree on how to implement the thing.

Another example is: A colleague recently wrote a post about how they’re working using agents, and it looks almost like an idealized life that many of us might want.

It’s like: I wake up in the morning, I think about the research that I want, I tell five different Claudes to do it, then I go for a run. Then I come back from the run, and I look at the results, and then I ask two other Claudes to study the results, figure out which direction is best and do that. Then I go for a walk, and then I come back.

It just looks like this really fun existence where they have completely upended how work works for them. They’re much more effective, but also they’re now spending most of their time on the actual hard part, which is figuring out: What do we use our human agency to do?

My colleagues are working really hard to figure out anything that isn’t the special kind of genius and creativity of being a person. How do I get the A.I. system to do it for me? Because it probably can if I ask him the right way.

Are they much more effective? I mean this very seriously.

One of my biggest concerns about where we’re going here is that people have, I think, a mistaken theory of how the human mind operates. I always call it “The Matrix” theory of the human mind — everybody wants the little port in the back of your head that you just download information into.

But my experience being a reporter and doing this show for a long time is that human creativity and thinking and ideas are inextricably bound up in the labor of learning — the writing of first drafts.

I have producers on the show, and I could say to my producers before an interview with Jack Clark or an interview with someone else: Go read all the stuff. Go read the books. Give me a report — then I’ll walk into the room having read the report.

But I don’t find that works. I need to do all that reading, too, and then we talk about it, and we’re passing it back and forth.

I worry that what we’re doing with A.I. is a quite profound offloading of tasks that are laborious. It makes us feel very productive to be presented with eight research reports after our morning run. But, actually, what would be productive is doing the research.

There’s obviously some balance. I do have producers, and companies do have employees. But how do you know people are getting more productive, versus they’ve sent computers off on a huge amount of busy work, and they are now the bottleneck, and what they’re now going to spend all their time doing is absorbing B-plus level reports from an A.I. system that shortcuts the actual thinking and learning process that leads to real creativity?

Yes. I’d turn this back and say I think most people — at least this has been my experience — can do about two to four hours of genuinely useful creative work a day.

After that you are, in my experience, trying to do all the turn-your-brain-off schlep work that surrounds that work. I’ve found that I can just be spending those two to four hours a day on the actual creative hard work. And if I’ve got any of this schlep work, I increasingly delegate it to A.I. systems.

It does, though, mean that we are........

© The New York Times