Entering the trillion-agent economy (ft. Rohit Krishnan)

In this episode, I sit down with my friend Rohit Krishnan - writer of the Substack newsletter Strange Loop Canon - for a hands-on conversation about what it actually looks like to build with AI agents today. Between us we're burning through tens of billions of tokens a month - I hit nearly 100 million in a single day this week - and we share what we're each running on our own machines. We dig into the quirks and surprising power of tools like OpenClaw, Claude Code, and Cowork, debate why AI remains stubbornly bad at good writing, and zoom out to ask what a world of trillions of agents might actually look like — and what economic infrastructure it will need.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Azeem Azhar?

  On Wednesday, Rohit, I fell just shy of 100 million tokens.

• 0:06Speaker 01

  You know, it is intelligence on a tab. Computer used to be a job that people did. There used to be like room full of people who were computers. They would compute things. And now it's machine. The next up is analyst.

• 0:17Azeem Azhar?

  I think it's not unreasonable to think that we will have hundreds of millions and then billions and then tens of billions and hundreds of billions of agentic systems running around the internet as automated infrastructure in a few years' time. In that world, what do you think an agent meaningfully is?

• 0:35Speaker 01

  I called it homoagenticus. Suddenly, the world becomes one where agents are the ones that are interacting mostly with each other. Going to websites for us or like doing transactions for us using your credit card. All of these things are the things that agents need to be doing. Which means like when you have a trillion of them, we need different coordination guardrails for them to be able to do tasks for each other.

• 0:54Azeem Azhar?

  Agents will need some mechanism to exchange value between themselves, where it won't so much be about transaction costs and communication costs between employees, but it will be about security and verifiability costs.

• 1:08Speaker 01

  I get LLMs to review all of my essays and I throw away almost all of the comments that it gives me. I can see that it pushes it towards a mean. It's like, oh, you were too colloquial here, take it out. It's like, no, like without that, you know, it'll read like the back of a breakfast cereal.

• 1:23Azeem Azhar?

  Arminia actually moves house this weekend into a new Mac Mini. It can see my lights and I can turn them on as I walk to my studio from the house and a few other things.

• 1:34Speaker 01

  The hardest leaps for my wife was, oh, I don't need to think about what I need to ask it before I ask it. It's like, don't worry about it. Just talk to it like, you know, it's your analyst and it'll just do things for you.

• 1:44Azeem Azhar?

  Or your husband.

• 1:46Speaker 01

  Your husband.

• 1:49Speaker 01

  But it listens better than most husbands that I know, including myself. Well, like it at least remembers things.

• 2:02Azeem Azhar?

  Today, my guest is my friend Rohit Krishnan. Many of you may know him from Strange Loop Canon. It's where he writes his essays on Substack. It's very hard to describe what Strange Loop Canon is about. Let's just say, there's not a lot of canons in it, but it is quite strange. I absolutely love the essays. Business, technology, economics with rigor and clarity. Rohit, you're so distinctive. You're an engineer, you're an economist, you're a hedge fund operator, you're a builder and today, a coffee drinker. Now, we have both been experimenting extensively with AI and more recently with what people are calling AI agents. Thinking through practical questions like how do you get these damn things to do useful work?

• 2:44Azeem Azhar?

  Through to theoretical ones, like what would an agent economy look like? And of course, we've been playing around with open claw. We are going to try to cover all of that. So let's get started. Rohit, thank you for making your Friday morning available.

• 2:59Speaker 01

  My absolute pleasure. Thank you for having me. I mean, there's very few things that are more fun on a Friday morning than to do this.

• 3:05Azeem Azhar?

  Oh yeah, no, I've been looking forward to this. I've been building a lot using AI tools, you know, Claude Code and Redplate and other things. And you have as well. So what's on your screen right now? What are you working on?

• 3:19Speaker 01

  I have a Codex thing running right there. I have three screens. That is basically doing a music creation little thing. I just made it for me where I can play on my computer keyboard and actually play music. So that's number one. I have another thing running here, which is my polymarket prediction and foresight for each one. That is the daily prediction thing that I run for myself. I have two Telegrams running there, which is talking, one is talking to my open claw. One is talking to my own Telegram one for all my daily to-dos and all that kind of stuff, normal stuff that you might do. And I have a last one running here, which is a longstanding project that a hobby horse I have of like, I call it Horace.

• 4:00Speaker 01

  I wrote a post about it a while back, trying to see whether I can get AI to write better, which is, as we all know, the hardest thing for. An AI to do.

• 4:09Azeem Azhar?

  I love what you've got there. I have committed about three hundred thousand lines of code this year. It it has been really, really nonstop. And I was reflecting that my coding world has changed remarkably with Opus four point five, now Opus four point six. But also now with these OpenClaw agents. OpenClaw, correct me if I get this wrong, Robert, it it is a it's an agent orchestration framework that you can run locally on your machine or in the cloud. It has a persistent memory and tool use. And you can, it kind of works the way a science fiction software agent works.

• 4:49Azeem Azhar?

  And and people have gone crazy for it. So so you've you've been doing some open core work. Did I get the definition right, by the way?

• 4:56Speaker 01

  Yes, I mean, basically, OpenClaw, you know, it went through a bunch of naming changes. But basically, what OpenClaw does is that, like, you you have these L L M's that can do useful work. But useful work that you wanted to do on your computer is different to, like, if you can actually access it on your WhatsApp, on your Telegram, on your Slack, you know, Signal, whatever, or email, what have you. OpenClaw basically is a way that, like, it's a it's a each kind of built this to actually connect to all of these normal data sources that you would have, would need it to talk to, and you can make it run persistently in some environment, wherever you like, you know. On I'm running it on my laptop, you're running it on a Mac mini, people run it on the cloud, whatever you like.

• 5:41Speaker 01

  And then you just talk to it, just like you normally talk to an L L M, and it can take actions for you, like scheduling meetings or like like responding to messages, cleaning up your inbox, getting a daily to do list. I mean, all of these things that I do on a daily basis now. I mean, I have tried using to do lists, for example, is something that I've tried using for twenty years now. I have never ended up sticking with one because they all suck and like they make me do different things. And I like the longest I used them for is like four days. But this one, it works because I need to change zero habits, like, and it does the work that I should be doing, which is like read through my emails and tell me what I'm supposed to be doing today compared to the history. It's not a hard job, but it does require a little bit of grunt work that somebody needs to do. Well, until like three weeks ago, it was me.

• 6:27Azeem Azhar?

  Ah, right. So tell me about how you you work with your agent. Do you have a name for your agent?

• 6:33Speaker 01

  I don't. I haven't. The main reason is I haven't settled on. Um, cause like, I mean, right now I I call it Morpheus, cause it's it's the name that I used to use back in the day, but I don't know whether it's gonna stick or not. But I do want, I have it just because like, you know, it's on my Telegram, so like, I need I need to call it something, and it's on a bunch of my groups. So I have to say at something before I can call it. It's called Morpheus.

• 6:59Azeem Azhar?

  My agent uh has had a name change as well. So my agent was called Mini Arnold, because in the second Terminator, the Arnold Schwarzenegger character comes back to to protect humans and protect John Connor. And then Chantal, who's on my team, pointed out that we had agreed a year or two ago that we should, if we do have agents, we should call them R something, per the asympt rule, R Daniel Oliver. So Mini Arnold has been renamed R Mini Arnold, which is quite a mouthful, but I still call it R Mini Arnold. So when you're working with this agent Morpheus, what is the the type of thing that you you do when you interact with it? Just get get us to some specificity.

• 7:42Speaker 01

  Sure. So let me put it this way, like a year ago, two years ago, actually. The most used app on my phone would have been probably like WhatsApp or Slack, right? Like email a little bit, but like WhatsApp or Slack. Sometime over the last year, the most used app kind of became ChatGPT, cause I would use it for everything as a as an interface to ask questions, learns, etc. If I get five minutes standing in a grocery shop, I would be asking something to kind of respond and so on and so forth. Today, it's almost like I'm moving back to the Slack world, except that now ChatGPT is in there. So now I can ask it any question that I want about anything that I want. And this becomes the interface with which I kind of get against the world.

• 8:23Speaker 01

  So I do three kinds of things with it today, and I don't think this is the like height of what can be done. This is just what I feel happy doing today. Cause I've given it all my, I have a personal laptop here. I said like, this is your home, you know, you get to do whatever you want. You wanna mess up the file system. I don't care. Go have fun, right? Everything is backed up. So now I run it here for, one is daily tasks to do that kind of stuff. The second thing that I do is like email triage and inbox and responses and stuff like that. It occasionally gets something wrong, but like it's fine, right? I mean, it's not the end of the world. We, we can deal with these kinds of things. So figuring out which emails I should respond to, how deleting all the junky ones and run sort of regularly, that kind of stuff.

• 9:05Speaker 01

  The third thing that I use it for is like interacting with this machine to see whether I have stuff running. Like I have a piece of code running. Can you interact with it and tell me how that's going and like rerun it or, you know, get the terminal access and kind of talk to those kinds of stuff. So like the third one is probably the least personal assistenty, if that makes sense. Cause you're, even if you had a personal assistant, you're not going to tell them like, hey, go check my computer and see how that terminal code is running. Give me the response, then I'll tell you what to do and you go run. That doesn't happen. The first two are very personal assistant. It's like, it's like having a, you know, your, your own genius personal demon kind of go and do tasks on your behalf. And just like clean up your admin life so that you can kind of solve that.

• 9:47Azeem Azhar?

  How fast is it responding to you and how often are you ping it?

• 9:51Speaker 01

  Oh man, I don't even know. I'm, I ping it like regularly, maybe once an hour at least. Sometimes very regularly, if I have something immediate ongoing in terms of sort of stuff. Like anytime I'm walking around and I'm like, oh, I need to do that, I throw it into this. And then it kind of takes care of everything else that needs to happen at a stateful way on the backend.

• 10:11Azeem Azhar?

  This is your open call agent, Morpheus. Have you added any new skills to it? Cause there are loads of skills out there that you, you can add.

• 10:20Speaker 01

  I haven't actually, cause I haven't found the, because I've, I've hand rolled a few things, but I haven't really needed to. I found that like when I add skills, it is for some specific thing that I want to kind of rerun repeatedly. And what I find now is that it's all my ad hoc stuff, right? Cause if I know something I need to do repeatedly, then there's usually a workflow that I've built and kind of runs. Like the prediction one or like the daily, daily news forecast one, I already have an agent running for it. So I does, I don't need to do that in my open call. Open call is like. My personal butler, I mean, I get to ask Morpheus anything that I needed to do today. If it was something that I would do every day, then that becomes an ongoing thing that can be run somewhere else anyway.

• 11:04Speaker 01

  I don't need OpenClaw to do it.

• 11:06Azeem Azhar?

  So I tend to take the view when I'm using these tools that I want them to do the most useful things for me, not the trivial thing. Right. So a couple of the skills I've given Arminie Arnold are the Foundry library and the Reflection library. So these are self-learning skills. And so during the context discussion, I can say, learn this. And so it now has learned that I like executive summaries and bullet points when it was quite back to me. I like the date format, which is reverse date, Y Y M N B D before every piece. So if it's asked a task that is complicated, even if I don't specify a plan, it needs to plan and give me the plan. And then I've had some really, really surprising things. So I had it design effectively an email fortress,

• 11:50Azeem Azhar?

  like a protection fortress against prompt injection attacks. And it has these six or seven layers. It went off and found some academic papers and some red teaming data and built a test rig and we've got that in place. I gave it one of my apps, which was a multi-agent deliberation system. So you can put a question to a panel of agents with different personalities and they argue over 20 minutes. And it used it and it said, this could be very useful for the analysis I do for you, Azim. However, this is designed for a human. So can I just build it as an API-driven version for me? And I, you know, goes, I've got, and it did in four minutes, replace my work, which had taken me a day. Right. I gave it a credit card. It's too nervous to spend any money.

• 12:33Azeem Azhar?

  It's got a $50 prepaid credit card and, you know, it's still like, should I do that test page or not? And like you, so I don't do the email work, but I do do detailed ad hoc things with it. And I cannot understand why it is as good as it is. So one example, I got it to do a market report on Tropic. And I said, here is access to Prism. So Prism is the exponential view research backplane. And my OpenClaw agent, Arminie Arnold, went to Prism, downloaded.

• 13:11Azeem Azhar?

  Over five hundred novelties we have on Anthropic and wrote the best thing in ten thousand words that I have seen about Anthropic. Far better than what I get out of a G P T five point two pro deep research query or the Manus deep research query. And and I have no idea what's going on under the hood in open claw, but it was remarkable. It has kind of changed my way of working.

• 13:37Speaker 01

  The the one way that I think about this is that like with Codex or Claude code, if you use it and you you ask it to kind of go do a piece of task. It is remarkably good at being agentic and going and doing it.

• 13:49Azeem Azhar?

  Right.

• 13:49Speaker 01

  And in one way that I like open clause is like you now have a front end and a bunch of connections that you've given it. So like ultimately, like I do this all the time of, you know, I have a like various databases of different things that I have running, some for fun, some for work. And I would routinely just spin up a Codex instance because I use Codex quite a bit and say like, hey, go write me a report about this or tell me something about this. Go figure something out. And it'll run for, you know, twenty minutes, thirty minutes.

• 14:17Azeem Azhar?

  Can can I just check? So so this is like a little bit of a hack because for people who may not know, Codex is a code development interface that OpenAI runs. It's their version of Claude code. And you're getting it to do market research, essentially.

• 14:31Speaker 01

  I I can get everything, you know, one thing that Frontier Labs are really bad at, just like L L M's are naming things. So Claude code is not for code and Codex is not for code, right? Not just for code. They're amazing at code, but what they are is effectively like L L M's in a loop that works. So you can give it a task and it'll figure out some version. It'll create a plan. The plan will have twenty five steps. It'll go through like in other files. It'll go research the internet, whatever. Keep doing things in a loop until it gets an output that you actually want. Obviously, it works really well for code, but also it works remarkably well if it's not code, right? I mean, some like your report is not code, it's words, but the process to get to the report would require a bunch of iteration and these things are really good at doing it.

• 15:14Speaker 01

  So something that I do regularly is like in what in the folder, I'll right click the folder, open in terminal. Codex and then I can ask it any question that I want about stuff in the Codex. I did my taxes this way. I did like, I did my taxes last year this way as well. Cause I am not, I, I don't care enough and I don't want to spend that much time. It's like, just go, you go do it.

• 15:34Azeem Azhar?

  Can I ask you the, the in Claude Co-Work, which is the, the platform that Anthropic has that was spun out of Claude Code. Co-Work, I found incredibly powerful as well and I, I use that extensively and, and I was blown away by a, a report it did on some of my, my in- investments. But it has some tool use and it can call on tools. Can the code versions like Codex use tools as well for, for analytics?

• 16:01Speaker 01

  Yeah, it can do the same thing. Like it can, you can give it skills, you can give it tools, you can give it whatever and it kind of does it. What I find is that like Claude Co-Work is more of a, it's a better interface. Cause OpenAI is also pretty terrible at like products, as we know from their naming schema, right? I mean, like you just look at it, but they do roughly the same thing. Cause you can give them, you know, MCP access to kind of various things like play a right.

• 16:25Azeem Azhar?

  Just explain what MCP access is.

• 16:28Speaker 01

  If you have a piece of code running in a loop on your computer and you want it to talk to something, a different application. It needs to go talk to that application somehow. It needs to either use an API, which is an application programming interface. Like it, it has to talk to that API. It has to talk to that database somehow. MCP is basically just a model context protocol. It's just a wrapper that you put around that such that like the models know this is the protocol by which you should be talking to something. So you want to use Chrome, Google has just shipped web MCP such that like agents will now know how to use Chrome better. Rather than trying to act like human, you know, open it up, go to X, Y coordinate, click, which is kind of a terrible way to do things.

• 17:08Speaker 01

  But I think Claude Code, Codex, these things are really, really good at like using the tools of our given device. Cause that's how they're trained, right? I mean, they're trained on a computer. So there, anything you can do with a terminal, they should be able to do. And as we all know from every horror story, you can do anything you want from a terminal.

• 17:27Azeem Azhar?

  Indeed, the two of us, you know, we are standing on the edge. We might not be at the absolute bleeding edge. But I don't think we are in the, the laggards category of using these tools. So I think it's helpful for people who are listening to give a sense, you know, practically of. What you can now do with a tool like Claude Coop, which I think you can get for 25 bucks a month or something, or what I made it to do with my open core agent, Arminianal. I'll give you an example. I read a story in the financial press yesterday that said there's high dispersion in the stock market. Now, what does this mean? It means that the NASDAQ index is staying about flat, but stocks are moving 11 or 12% over time

• 18:12Azeem Azhar?

  because some are going up and some are going down. And that is a warning signal that it's a 99 percentile dispersion. And it may mean that there is a severe correction about to show up, not a bubble bursting, but a 15, 20, 10, 50, 60% likely, perhaps. I didn't really understand this idea. I'm talking with more sense now than I would have done yesterday. I just copied that and I threw that into Open Core and I said, go and make sense of this for me. Right, go and compare it to my portfolio and explain what's going on. Take your time, build a plan, spin up subagents because Open Core can spin up multiple subagents that go off and do other things, which is a better way of doing things.

• 18:54Azeem Azhar?

  And it came back in 20 minutes and it was a really, really good report. It had gone out and got historical data, they've got stock data and assembled this for me. And you know, like you, I was leaving in a car journey, fired that off, finished the journey. There's a report in place, which I've now acted on, right? I've now acted on the basis of that. So that's a kind of practical, very hard thing that might have taken me a day and I would have never got around to it. How about you? What's a great example for you?

• 19:25Speaker 01

  There's this essay that I wrote, I think, three years ago. It basically said like, computer used to be a job that people did. Like when we went to the moon first, there used to be like room full of people who were computers. They would compute things and now it's machine that like you just, you do everything on the machine. And I said like, the next up is analyst. That was my kind of provocation in that, in that particular essay. This was sometime in 2023 and I feel like today we kind of have that. It's remarkable the speed, like anything that I want to analyze about any subject, any topic, anywhere is effectively at my fingertips. The only thing that remains is for me to ask, like that is the thing that remains, right? So.

• 20:06Speaker 01

  I mean, it's really hard to explain, you know, where would you use it? Because I use it like this is my this is the prism with which I interface with the every part of the world. Anything that I don't understand, I ask. Anytime I see something similar to that, that I kind of want to find out, I go and either search or create a report or get it to kind of do some work for me. Like if I'm trying to figure out, hey, I want to do, you know, this particular project. Which book should I purchase or buy? Can you kind of compare and contrast? These are simple ones or more complex. If there is any analytic work that I kind of need to do, either as work or like for myself, financial analysis. It's not like the components are, you know, quantum physics. I mean, these are solvable problems.

• 20:50Speaker 01

  And there is a ten, twelve, twenty step process that I would have taken if I had enough time and energy to kind of go off and solve that problem. But I don't have the time, which means most of them would have not gotten done. Except now, I I can just do it, right? I mean, I can spin up a report. It's amazing that like, one way to think about it is that you have an analyst at your disposal and you can spin up a report about anything in the world in twenty minutes. Will this be the best report in the world? Maybe not, but like it's a extremely valid research, ninetieth percentile report on any subject and topic that you want in twenty minutes. I built one for paleontology recently, just because like my son and I were talking about it. And I had this idea that like, when climate change kept happening in like history,

• 21:32Speaker 01

  what is the relationship between the variants in climate changing? And like the number of taxa, the species that actually existed. Like, I'm not a paleontologist, right? I I love the subject, but like, I can now solve that. I have I have a project running to actually solve that problem for me.

• 21:50Azeem Azhar?

  It is amazing, and I would just wanted to give a flavor to people who perhaps have not had a chance to get a handle on Claude cowork or on Codex or an OpenCall agent, how it changes your behavior. I'm going to give you a little bit of a shocking statistic. It may not shock you, but my token usage has increased a little bit since I started using OpenCall. So just again, for people to understand, you know, a token is about three quarters of a word. If you go Back and forth typically to the chat G P T you might use four or five thousand tokens in a in a single instance if you use a deep research query it might be yeah tens of thousands or hundreds of thousands of tokens.

• 22:27Speaker 01

  Yeah.

• 22:28Azeem Azhar?

  But but personally, I was probably using a million tokens a day with various workflows that are running up until open call. And on Wednesday, Rohit, I fell just shy of one hundred million tokens personal usage per day. And in fact, this week I averaged about eighty billion tokens per day. And each one of those is generating value because I know how much I'm paying, I'm paying yeah tens of bucks a day, so it's thousands a month. And I'm getting that value back and I can see it. And it is unended, it's what spurred me to write this update to the boom or bubble question saying, listen, demand is really off the charts and it's going to continue to go off charts.

• 23:12Azeem Azhar?

  So a hundred million tokens a day is my new baseline.

• 23:15Speaker 01

  It is crazy that like I did this analysis, I think, last year. So I built a tool to kind of monitor all my agents usage on my computer. Again, one of those things you can do because you have a question, you can build a tool for it. So I did it and I built it in Rust for those of you who are kind of listening, it's not a language that I know how to code in. But I was like, now I can because I don't need to know because I know Python and I know C plus plus, but I don't know Rust, I can do it. Anyway, so I built this tool and I learnt that in the last quarter, the Q four of twenty twenty five, I spent about seventeen billion tokens, give or take, in Q four. Then I realised in month one, in January of this year, I spent fifty billion. It the spike is in and once you see the trend line here.

• 23:57Speaker 01

  And the thing is, I distinctly remember because it was only like two years ago, when you would like do token counting. When you're sending a query, when you would be like, oh, man, it's like ten thousand tokens, should I kind of send it? It was like two, three years ago, it wasn't that far back. But now it's like it's at fifty billion last month. You know, it is intelligence on a tap that you can kind of turn on and you can do anything you want with it. I mean, why would I ever do a research report on like marine mammal, you know, fossil convergence? Like, there is no logical reason for me to do that, except that now I can, because I have a piece of curiosity that I can actually solve for. And that applies both for work as well as for like life, right? I mean, you can solve anything you like.

• 24:37Azeem Azhar?

  Note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show. I thought Horace, your writing tool was called Walter.

• 24:50Speaker 01

  No, no. So I have a, I have, I have two. Walter was my first attempt at trying to do something that actually works properly. Now I have a different one called Horace. And the purpose behind it is basically to say like, can you measure literary cadence across different texts, learn from AI and be able to sort of test it? It's a hobby horse, right? Because we are both writers. And one of the things, one of the weird things about the last few years as being writers is that there is a text generation machine that someone created, right? Like it is the world's best text generation machine. Turns out what it's worst at is actually generating good text.

• 25:31Speaker 01

  But what it is amazing at is like generating code, generating images, even videos. It's getting really good. But if I ask it to write like a four paragraph, like essay even, it's not very, it's not engaging. It's not fun to read and it just annoyed me. So that is why I tried to figure out, can you get the cadences of these like writings over a period of time? And then try to capture from it some lessons, train a model against it and then be able to deploy it so that I can test various pieces of writing. Saying, is this literary enough? Is this well written enough? Does it have the right cadence enough, et cetera? Basically just to get AI to say like what AI generated writing is, it's slop.

• 26:11Azeem Azhar?

  So Nathan Lambert, who runs Interconnected, wrote this, this great essay about AI writing. He said like AI writing is distinctly mid. And that has been my experience as well, right? It is in the middle of the statistical distribution. It doesn't just delve, i- if this is, and then that's quite a lot. There are lots of these signature tells that uses a lot of Latinate words. And, and so like you, I have been building a tool just as an experiment and mine is called Broca. So Broca is the language center of the brain. And the ambition of Broca is to be able to get one of these things to write well. So I think we have a chance to maybe. Compare notes, so.

• 26:51Speaker 01

  Yes, we should absolutely do that.

• 26:53Azeem Azhar?

  So let's.

• 26:53Speaker 01

  Rock is a fantastic name, by the way. It is a brilliant, brilliant name.

• 26:56Azeem Azhar?

  I'll tell you where all of the names of my products come from. They come from scientific themes that are connected to the thing that we're trying to do. So there's another one that is called Prism and Prism is the research backbone. Cause you kind of look through the prism, the research backbone. There's another one called Scintilla, which was the early signals detection. The trouble is I've built so many, I actually can't remember what any of them, any of them.

• 27:21Speaker 01

  See, when the names are too cool, you can't remember all of them. But you know what's really funny? I have a daily running thing that generates predictions that I read sort of as a different lens to watch news.

• 27:33Azeem Azhar?

  Right.

• 27:33Speaker 01

  This is the only one where I let the AI name itself and change it over time.

• 27:37Azeem Azhar?

  Yeah.

• 27:38Speaker 01

  Guess what the name was? Foresight Forge. It is the most obvious of the slop names. I kind of love it because like, you know, I give you full freedom and control and you can revise it as many times as you want. And you still land on slop. Like that tells you everything you need to know.

• 27:53Azeem Azhar?

  Yeah, absolutely. And I think if anyone's been using Replit, which is something that I use from time to time. The names that Replit automatically gives the software are exactly that sort of structure, you know, like finance boundary. And they have to alliterate, they have to be two words. So, so let's get into Horace, which is your, your new version. How well is it working?

• 28:12Speaker 01

  It's working okay. Here's a discovery that I've made before I get into how well it's working. What I found is that my hypothesis was if I take like essays from, you know, or Charlie Strauss's short stories. And if I get AI to write a short story in similar fashion, how well will it be able to distinguish? What is kind of interesting is that if you get the best AIs to actually write short stories, it mimics the cadences really well. In the sense that it has learned some underlying structure. It's just that it's, it's like a child putting a Lego set together, you know? They don't care about the right colors or they might not care about the right shapes in some ways. They just kind of built it and say like, here's a whatever, you know? Here's a spaceship or here's a castle, here's a dinosaur.

• 28:54Speaker 01

  And you kind of have to look at that and go like, if somebody wanted to really build a dinosaur, they would have had to change. Change a bunch of these things to actually make it look good. So Horus works okay today. You can extract some per token metrics kind of works well. It's not ready for prime time yet, but it is something that I'm intrigued by because like if it works, it's a good counterpoint to something like a Pangram Labs, right? Where you can take a piece of writing and say like, oh, if I do these things, it actually will sound good, not just like non-slo, but will actually sound good. One maybe snippet of what was interesting is like, I first tried it with pure token and sentence based metrics and that all worked super well. Then I thought like, just like you said, answer to everything in life is more tokens. So why don't I take a bunch of short stories, essays, novels, poems that I love?

• 29:40Speaker 01

  And just extract the main sentiment of that, give it to AI to rewrite it. And then I swap them. So I can say, here is the sentence, this is what good looks like, this is what bad looks like. And now instantly it was like a big improvement.

• 29:56Azeem Azhar?

  If you think about how, you know, a transformer works, it's next token prediction. It is able, it's clearly able to divine some lower level structure, some substructure that we don't see, whether it's because of, you know, the high dimensionality of the space or some other structure it's finding. But it's still a statistical game. And one of the things about writing is that, of course, I have a style of writing and you can recognize my writing if you compared an essay from Tolstoy to mine, well, his is better anyway, but you'd know which one is mine. And, and so, and the approach I've taken has been to take a bunch of my writing, unfortunately, I've got hundreds of thousands of words. And view the, the writing as a kind of fractal process. There are lots of layers.

• 30:40Azeem Azhar?

  There's paragraph length, there's threads that run through, there's sentence length, there's word cadence, there is lexical etymology. Do I use Latinate or Germanic words? And I use some really traditional natural language processing tools, sentence parser, fragmenters, tone, concept, argumentation style, to build a fingerprint of my writing. And then I use that writing to, that those parameters to pass as you must or you shouldn't to the L L M. And effectively we've gone through, I have a little bit of a testing rig. So it writes, it writes an essay, I look at it, I go. I mark it up, it goes back through and it figures out how to adjust the weight.

• 31:26Azeem Azhar?

  And how good is it? It's a B plus at three paragraphs right now. You know, it's been a fun, fun process. It is teaching me a lot about my own writing, frankly, which, which is quite interesting. I didn't know, for example, that I use eighty percent Germanic root words rather than Latinate words. Most L L M's are about sixty percent Latinate, it turns out.

• 31:49Speaker 01

  That is a fascinating. I mean, the L L M's have these weird like peccadilloes, right? And different ones have different sort of interests. Like, for example, I get L L M's to review all of my essays and I throw away almost all of the comments that it gives me. Those old comments, I can see that it pushes it towards a mean. It's like, oh, you were too colloquial here, take it out. It's like, no, it's meant to be like without that, you know, it'll read like the back of a breakfast cereal. Who wants to read that? It has to be interesting for somebody to actually understand it. Or I think writing weirdly is like the hardest eval for all of them. Because for math, for science, for research, you can almost still do like L L M as a judge and get them to figure out if it's going well.

• 32:33Speaker 01

  For writing, you kind of still have to do it because they have no taste yet.

• 32:36Azeem Azhar?

  I don't think they have taste, but I think the fact that you identified as being so hard is, is also true about writing well in, in general. So all of us write and, and, and, but most, most people don't write well enough. And I think if you go off and you look at really great writers, whether it's looking at a George Orwell or you look at literary fiction and you, you think about what goes on in the construction of that, that world model, that mental model of the world. What gets disclosed, what doesn't get disclosed? When does your use a sentence that runs over four pages? When they use a sentence that is one word long? These are deliberate choices. This is not page rank with a teleport every tenth link.

• 33:22Azeem Azhar?

  And that model is built in the head of the writer. And I think it's extremely hard. I think that's why you'll often find places which have got exceptionally brilliant people, like an investment bank, the type you used to work in or, or a top consultancy. And the written output is is terrible, which is why these consultancies hire so so many writers to try to polish the work often. So I think it is a testament of of, I mean, what is it? Is it that when we write, we are combining our assessment of empirical space that we are considering, our experience and some measure of interiority that is not captured anywhere else? Is is that what's going on?

• 34:04Speaker 01

  I suspect so. I think one way that I've been thinking about it is that like L L M's are incredibly good at learning patterns. You throw any data at it, it can identify hidden patterns in there if it has seen it in the training data in some sort. And whether that's pre-training, post-training, R L, whatever, it kind of doesn't matter. It's really good at that. But every test that I have done kind of shows that like it is very happy to identify a set of those patterns. And then once it identifies, it just kind of applies them, right? I mean, there is no there is no interiority necessarily in the way that you say. Now, it's really hard to talk about it because anytime you understand a concept well enough that you can nail it down to like some specifics, obviously that can then go into the training data again and like that loop continues.

• 34:49Speaker 01

  So it's a little bit like catching water. But at the same time, I do feel that like when I'm reading, you know, if it's Orwell or like Charlie Strauss or whomever, if I'm reading an author that I love, there is a sense of who the author is in the book that I'm reading. It is not just a construction that usually stands completely separate to who the author is. There is a personality that kind of emerges and L L M's are still kind of flat in the amount of personality that they actually are able to kind of push out and emit. They have such a strong tendency to come back to the median, the main patterns that they have learned that these kind of, I don't know, nonlinear, non-easily interpretable patterns that might exist in writing is the hardest one for them to stump out.

• 35:33Speaker 01

  Because the the point of like, can you get this point across? They're very good at. But can you get this point across like a Zim? They're not so good because they're like now they have to kind of mimic a Zim and maybe they can mimic you in the surface, but like there's a there's a lot going on, right? The essay that you write on a Friday is different to the essay that you would write on a Thursday.

• 35:52Azeem Azhar?

  Absolutely. I mean, and I, we know this because we test it all the time. And it's like, well, yeah, that's our one framework that I used 10 years ago, it's suitable then, you don't bring it out now. I want to move on to something else in a second, but I'll give you one thing that I found quite helpful. So this is a tool that I built called Mirror. And Mirror is back to you. So it reflects a process that I go through when I'm writing myself, which is I ask myself questions or my team asks me questions. And what I do with Mirror is I will come up with like the thing I want to write about, the argument I want to make.

• 36:22Speaker 01

  Right.

• 36:23Azeem Azhar?

  And I will then connect it to one of our data stores. So maybe it gets connected to Prism, so it can look at all of our research. And then there is an agent that plays the role of a good editor and is entitled to ask me up to six questions.

• 36:37Speaker 01

  Right.

• 36:38Azeem Azhar?

  And it's quite tough. It asks me these questions and what it comes out with is a structure. So not an essay, a, an outline structure, which I can then look at and use as a critique of, is that the structure that I, that I want? And, and, and I think what I said to you earlier, that like writing is very fractal. And can you get the L L M? I think it's very difficult. In the same way as using an L L M for code, if you just try to get the L L M to write a complex bit of software, it will fail because it sticks everything together. If you break the task apart, architecture, front end, back end, database, security, it's going to have much higher degree of success. And I think the same is true with that human writing where I separate out the structural component from, you know, the other, the other components.

• 37:23Azeem Azhar?

  And, and that has been one little insight I had in the last few months.

• 37:28Speaker 01

  I think that's a very interesting one. I feel like, I mean, I do use them to ask like, is this logical? And that is something that they get there better at answering than like, is it written well, if that makes sense? But the thing that you said about them being fractal is super important because that's something I think about quite a lot. Because like L L M's are now are very, they have gotten good at like a sentence level. If you ask them to generate a sentence, they're, they're, you know, it's still not always, but 50 50, you will get some good sentences. What you don't get is that sentence in a paragraph. They're not there yet. Or that what you don't get is that paragraph in like a section or that section in an essay. Or, you know, like as you go zoom out more and more, they get worse and worse. Now, the funny thing about code is that like,

• 38:13Speaker 01

  I mean, think about the amount of effort that has gone into getting them to write code, right? Like, they've read every single piece of code that is existing out there. They self-generate the data and learn again from that code. We are spending, you know, billions, tens of billions of dollars a year trying to get them to write code better. So it kind of is like, yes, they, it, it worked because of the amount of effort that we have put into it. There is no similar effort that we have put into writing because like we can't. Because you need something to judge it before it can go, right?

• 38:43Azeem Azhar?

  Yeah, that verifiability, but the code breakthrough has been pretty amazing. If we'd had this conversation in September, it would be completely different to December, and it certainly would be different to today. You and I have got these, this, this, at least one agent that can spawn multiple sub-agents. I think it's not unreasonable to think that we will have hundreds of millions and then billions, and then tens of billions and hundreds of billions of agentic systems running around the internet and this as automated infrastructure in a few years' time. In that world, what do you think an agent meaningfully is, right? What is the, the, the meaty definition of an agent?

• 39:26Speaker 01

  Really good question. This is something I'm spending a lot of time working on. And I'm like, actually, I'm writing a paper that'll come out hopefully at some point soon about agentic economy and why it might take to coordinate, et cetera. So one starting point is to say that like there are eight billion humans on the planet, and if eight billion humans actually start using agents in any meaningful sense, there is no reason we shouldn't get like a trillion agents. This sounded fanciful even like a quarter ago or a year ago, but like I already have like twenty of them run. And, you know, I'm barely kind of touching the surface of this.

• 39:56Azeem Azhar?

  Can I challenge you there? You've got twenty agents running. Are you got twenty agents? Or have you got one agent and a bunch of chrome jobs?

• 40:01Speaker 01

  No, so many of them are like full-fledged agents that are actually going off and doing things individually. I think, see, the second part is important because like, what is an agent? This is the kind of definitional question, right? Fundamentally, right now, what an agent is, is some version of a persistent L L M that has its context changing continuously, relatively autonomously.

• 40:22Speaker 01

  It's like ultimately like your open claw instance, you're still asking a question. The question along with a bunch of context is going into Claude opus four point six. That is doing some analysis and sending you back some information. Like the fundamental unit is still Claude four point six into it. Some question comes and then it says, oh, you know, Azim asked a hard question. I might spin up like six more to kind of go do different things come back. But it's context management all the way through. All of us have these agents kind of representing us in the world. Suddenly the world becomes one where agents are the ones that are interacting mostly with each other. And if agents are the, or you know, we are in the mix. Don't get me wrong. But like agents are the ones that are doing the jobs for us or interacting with us. Going to websites for us or like doing transactions for us, using your credit card. All of these things are the things that agents need to be doing.

• 41:06Speaker 01

  Suddenly, as more and more autonomy comes towards the agents, we have to figure out what are the kind of institutions we need. Institutions of the word I use for agents such that they are able to do the task that you gave them to do. Cause agents, just like you said, are weird, right? I called it homoagenticus in the sense that like it is a new genus. It does diff- things differently. They would much rather build versus buy. They're scared of spending cash at this moment. They don't really like trade. Like there is these weird things that show up in their behavior. Which means like when you have a trillion of them, we need different coordination guardrails for them to be able to do tasks for each other. But if you're having this conversation again in like a year, two year, three years timeframe, like that twenty will become two hundred. Cause we would just have a bunch of stuff that like they would be doing in the background.

• 41:51Speaker 01

  Things that would cost a thousand dollars a day today would cost like a dollar a day, like two years from now. And you can just, that's, that scarcity is gone, right? And then you can suddenly go up and like for as much as my coffee costs, you can kind of do all sorts of incredibly complicated analyses. That is the speed with which this changes.

• 42:09Azeem Azhar?

  That's a fantastically articulate picture, which I agree with so much of it. We may have one or a handful of agents that are the primary agent. I called it the AI chief of staff because I don't actually want to remember all of these tools that I have. I, I want the performance. I want someone else to log it. And I don't want to be poking around, well, which tool do I pull out at this moment? So there is a sense that perhaps that's what the open claw agent, or this is exactly a great position for Apple to find itself in. Which would be, we do, we, we can. Own that that agent interface and that trust interface. And then there is an enormous number of agents out there. And I think you're absolutely right.

• 42:54Azeem Azhar?

  You do need to have some transaction mechanism. A transaction mechanism is required because there are still resources. And those resources may be latency, right? Responding one millisecond is more expensive than a thousand milliseconds. Compute, energy, there will be things that are non-rivalrous that we've made rival because of law. So in terms of types of information and data.

• 43:19Speaker 01

  Right.

• 43:19Azeem Azhar?

  So so agents will need some mechanism to exchange value between themselves. And we will replicate, it would be my bet, some of the features. We'll replicate a Cozy and boundary of where it won't so much be about transaction costs and communication costs between employees. But it will be about security and verifiability costs. And I will want internal facing agents that are less sandboxed than the ones that actually interface to the outside world. And then the outside world ones will be interfacing on much tighter contracts and be spending some kind of resource or money in order to get things done.

• 44:01Speaker 01

  You need some way for the agents to kind of discuss, negotiate, like talk to each other such that they can actually interact. You need some version of identity. You need some version of verifiability that like they did something. You need a unit of exchange. In fact, one of the essays that I wrote, like, I think, two, three weeks ago with Alex Emes was on will we need money in an agentic economy? Because we were trying to show that theoretically agents are like humans in the sense that they can talk to each other. But if you want them to do stuff together, they need some shared medium of exchange so that they can, you know, it's like the hierarchy and price signal thing. Like you need something so that they don't need to solve everything from first principles every time, which is not a sensible way to kind of run anything. We're going to start seeing that. We'll start seeing that this year. That's my prediction.

• 44:46Azeem Azhar?

  It might actually be the use of a certain class of crypto, right? You know, the way of exchanging back and forth. Because I'm not sure the the scale, right? The quantum of some of these transactions. I mean, it's really important that there is a there is some value astride. Otherwise, you can't astride value. And what will that be? You know, maybe PayPal and Stripe will deliver rails for payments with fractional sense, fracto sense, we could call them. But I, like you, agree that there will, there's quite likely to need to be this means and and mechanism of exchange. So it's kind of the world that we already know.

• 45:23Speaker 01

  I mean, plausibly, that is one way that it could be. Like, there is no reason that it should be so de novo new that like, because we know how the economies work. We know how human economies work. And the human economy that we are having this conversation in in 2026 has evolved over like several hundred years, thousands of years, where we have tested different things and we have stuck with certain things and they keep coming back. Like, medium effects change, keeps coming back. Identity keeps coming back. Like, various checks and balances keeps coming back. The separation of fiscal versus monetary policy, if you want to kind of go all the way up, keeps coming back. Like, every time something, like, these are, in a weird way, like, economic invariance, just like there are physical invariance. So I would not be surprised at all to, I would expect to see something similar show

• 46:06Speaker 01

  up in the agentic economy as well, where we have similar invariance. Now, will it look like a greenback? Probably not. Will it have the picture of the queen on it? Maybe.

• 46:15Azeem Azhar?

  We have pictures of the king now.

• 46:17Speaker 01

  Pictures of the king now. I gotta get the new pants. But like, these are, these are things that will change over a period of time, but like, ultimately, a function that they solve will not necessarily shift. Because ultimately, that is kind of what we'll still need to solve for. Like, they do need to solve same problems that we need to solve.

• 46:33Azeem Azhar?

  A lot of questions from people about how should they get started with OpenClaw and have them use it securely. So I don't want to pretend it's easy. It's easy to do the one-click download and get something running. And they're off, but there is a world of pain, config files, dot N variables, and so on. The, but, you know, one of the easier ways, especially if you're nervous about putting it on your home network, is you can go and get a VPS, which is a virtual server. A company like Hostinja, Host Inga and Hosting or Digital Ocean both offer a VPS. It's seven to fifteen dollars a month where you could get get started.

• 47:16Azeem Azhar?

  And then you can decide how many resources you actually want to give that that agent. Do you want to give it your email logins? Do you want to give it your logins to your PayPal or or not? And, you know, perhaps don't do that at that point. So that's kind of quite a quick way to get started. You sync it up with a Slack channel or a WhatsApp channel. There are there are instructions, although it's pretty easy if you download the Mac in store as well to do it that way around. Let's talk about securing it. What do you do about securing your agents?

• 47:44Speaker 01

  I just, I I am more free rolling in this case. I live YOLO because I I gave it access and now it's like, let's see how much you can screw it up. It's fine, but I don't recommend that to most folks. I would I would highly, highly say. I mean, my basic philosophy is as follows. You need like L L M's basically work on the basis of context that it it is given. If somebody poisons the context, like L L M's will not know what is poison versus what is not very easily. As the models get smarter, they get better at distinguishing this, but it's still really, really hard. However, on the other side, I want to talk to it on my Telegram. I wanted to be able to access my emails. I wanted to read through all of my previous history and my documents and stuff.

• 48:27Speaker 01

  If I wanted to do that without detracting from functionality, it's really hard to put a security layer on top of it. So step one that I would say is like connected to something where you're okay with like it not sort of it going off into the world and like screwing up something, right? So the easiest way that like I have on boarded people is to say like, choose a folder on your computer, download Claude Code or Codex, open it in terminal. I know it looks scary, black, whatever if you're, but it's not. It is the easiest thing in the world to use because if you open up a terminal and type, your interface is literally just typing. It just looks like it was built in the nineteen nineties because it was, but like it's still a much easier way in some ways than to kind of install OpenClaw. Because installing OpenClaw, it's again, it's not that hard.

• 49:12Speaker 01

  You've got to install something on your thing, but you've got to connect it to Telegram. Then I would say like, just use it for a day so that you can just talk to it. Get used to that interaction method. And then over a period of time, you can say like, oh, I wanted to connect to my Google Calendar as well. Okay, I need to download GOG and you check and somebody will tell you NPM installs beer. And like, it'll start working, et cetera, and then you can kind of do it. Then you can say like, okay, maybe I should, I'm doing it long enough. I need a virtual server. I need to get Hetsner or Digital Ocean or whatever. Then I can, like, I would say like, go slow because people get scared because computer installation instructions, if you're non-technical, look like gibberish, because they are gibberish, right? Like, none of these words mean anything. But you can just stop worrying about it.

• 49:54Speaker 01

  Think of them as like a magical incantation, copy paste it in and go. That's at least my advice.

• 49:59Azeem Azhar?

  Yeah, I, I think that there's some sense also in getting to use Claude Code or CoWork in, in the first instance, so you get a sense of what the agent experience is like before you jump into, into open clause. So, you know, get, get the premium version, get Claude Code and CoWork. Try some experiments, build up your ability. The trouble with the BPS hosted service is that it doesn't see any of the stuff inside your home. And so our Mini Arnold is, is on a Mac. Our Mini Arnold actually moves house this weekend into a new Mac Mini, which has got much more RAM because it needs to run a lot for subagents and it kept hitting memory pressure and PS kills that were, were killing processes.

• 50:44Azeem Azhar?

  But as a consequence, it can see my lights, it can see, and I can turn them on as I walk to my studio from the house and, and a few other things, uh, which I can't do if it's in a, in a VPS. I have a setup here, which is probably a little bit more complex than where the average home. You know, I have got a couple of firewalls and a fiber line and I can isolate things. And so that may not be your experience, but I would suggest for anybody who wants to do this, start with Claude Code or CoWork or Codex, your recommendation, to get a sense of what you can do agentically. Before you fall down the open floor.

• 51:21Speaker 01

  Yeah, can I just say like one thing that I have found is that I I set up my wife with this late last year at some point, late last year. And I think the hardest leap for her to make, it it took a week, but it the hardest leap was for her to be like, oh, I don't need to think about what I need to ask it before I ask it. Because it's very common, like, how should I phrase X, Y, Z? Should I do this or it's like, don't worry about it. Just talk to it like, you know, it's your analyst and it'll just do things for you. It took up.

• 51:49Azeem Azhar?

  Or your husband.

• 51:52Speaker 01

  Your husband.

• 51:54Speaker 01

  But it it listens better than most husbands that I know, including myself. Well, like, it at least remembers things.

• 52:00Azeem Azhar?

  Rohit, thank you so much for joining. Your Substack is Strange Loop Canon. I cannot recommend it highly enough. And of course, for everyone else, if you're not subscribed to Exponential View, please subscribe to Exponential View. We're on Substack, we're on YouTube and wherever you get your podcasts. Rohit, thank you very much.

• 52:22Speaker 01

  Thanks, Azim. This is awesome.

• 52:24Azeem Azhar?

  Cheers. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now and I'll catch you next time.