Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop".

Jane: The paper was written by Igor Itkin from.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we're looking at a paper that made me laugh out loud when I first saw the title, because it's so honest about what it's trying to do. It's called "Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop."

Jane: And honestly, Tom, that title is perfect. We've all seen those papers where people simulate a thousand AI agents and it costs them a fortune in API calls. This paper is basically saying, hey, what if we didn't have to do that?

Tom: Exactly. And I love that the author, Igor Itkin, is an independent researcher. No big lab, no massive compute budget. Just a laptop and a clever idea.

Jane: So let me try to explain the core idea in simple terms. Imagine you want to study how a whole city behaves — how people spread opinions, how they react to an economic shock. You could simulate every single person as a full AI agent, which is expensive and slow. Or you could ask a small number of AI agents how they'd respond in a few different situations, learn the pattern, and then replace every person with a tiny mathematical model that follows that pattern.

Tom: Right, and the paper calls those tiny models "surrogates." Instead of a full large language model making decisions for each agent, you have a simple formula with maybe a dozen numbers in it. And the wild part is that for many questions — like does a society show a Phillips curve, or does an epidemic spread — those simple surrogates give you the same answer.

Jane: But here's the catch, and this is what I find really clever. It doesn't always work. The paper spends a lot of time figuring out when it works and when it doesn't. And the deciding factor is what each agent can see.

Tom: Yeah, that's the perception thing. If every agent sees the same global signal — like a national inflation rate — then averaging works great. But if each agent only sees their own neighborhood, or their own private information, then a simple average model just breaks.

Jane: And the paper even has a fancy diagram for this. They call it the interaction order times memory taxonomy. Global signals, community signals, local signals. Short memory, long memory. Each combination predicts whether your cheap surrogate will work or fail.

Tom: I love that they tested this blind, too. They pre-registered their predictions before running the experiments, so they couldn't cheat. And mostly the predictions held up.

Jane: Which is more than most papers can say. So the big picture here is that we might not need to spend thousands of dollars to study AI societies. We can spend a few dollars, run the whole thing on a laptop, and still learn something real.

Tom: And that opens the door for researchers who don't have big budgets. Independent folks like the author himself. I'm excited to dig into the actual results, because there's a lot more here than just the cost savings.

Jane: Oh, definitely. There's a whole section about what the macroscopic numbers actually measure, and whether they're real or just accounting tricks. That's coming up next.

Summary: Tom: So Jane, we've established that this paper, "Poor Man's Agentic Modeling," is about replacing expensive AI agents with cheap mathematical surrogates. But the summary section has some real surprises in it.

Jane: It does. The first thing that hit me was the EconAgent result. That's a simulation of a macroeconomy where each household is an AI agent. The original paper claimed it reproduced Okun's law — that's the relationship between unemployment and GDP — with a correlation of negative zero point nine one eight.

Tom: And this paper basically says, yeah, but that's because Okun's law in their simulation is an accounting identity. It's like measuring your height with a ruler that's also your height. Of course they correlate.

Jane: Exactly. They showed that a completely random, behavior-free policy — just flipping a coin for whether to work — already gives you an Okun correlation of negative zero point nine nine eight. So reproducing Okun's law proves nothing about the agents being smart.

Tom: But then the Phillips curve — that's the inverse relationship between unemployment and inflation — that one is real. The original paper reported negative zero point six one nine, and this paper's surrogate, fitted from real AI decisions, got negative zero point five six nine. Within noise of the target.

Jane: And here's the beautiful part. The surrogate has this one parameter that controls whether people work more when prices go up. That single parameter, fitted from maybe a few hundred AI decisions, is what produces the entire Phillips curve. It's like finding the one lever that makes the whole machine move.

Tom: So they're not just reproducing the numbers. They're identifying the mechanism. And then they do this really clever ablation study where they ask: is it the reasoning that causes the Phillips curve, or is it the wording of the prompt?

Jane: Right, that two times two experiment. They crossed whether the AI is asked to reason step-by-step with whether the inflation signal is described in plain or amplified language. And the result was stark. Without reasoning, the Phillips correlation was weak and even flipped sign depending on wording. With reasoning, it was strongly negative under both wordings.

Tom: So the reasoning step — actually thinking through the decision — is what creates the macroeconomic law. Not the phrasing. That's a profound finding, because it suggests that the way we prompt agents changes what macroscopic behavior emerges.

Jane: And it means the surrogate isn't just a cost-saving trick. It's an instrument. You can use it to measure which microscopic ingredient produces a macroscopic phenomenon. That's the real contribution here, I think.

Tom: There's also this whole section on the De Marzo consensus game, where agents adopt the majority opinion. The paper shows that the critical group size — where consensus breaks down — is actually a perception threshold, not a thermodynamic one.

Jane: Meaning it's about how well the agent can read a weak majority in a long list of opinions, not about some fundamental physics of consensus. And they proved that models with perfect counting ability never lose consensus, no matter how many agents you add.

Tom: That's a really clean result. And it explains why more capable models have larger critical group sizes. They're just better at reading the room.

Jane: So the summary is dense, but the through-line is clear: cheap surrogates work when the perception structure allows it, and they fail when it doesn't. And when they work, they tell you something about the mechanism.

Tom: And I think the next segment is going to get into the improvements and what this means for actually building these simulations. Let's take a short break and come right back.

Improvements: Tom: Welcome back. We're still on "Poor Man's Agentic Modeling," and I want to bring in our guests now, because the improvements section of this paper has some serious engineering implications.

Jane: Absolutely. Let me hand it over to Lu first, because I think you had some thoughts on the memory kernel measurement.

Lu: Thanks, Jane. Yeah, the memory kernel result was the one that really caught my eye. The paper measures how an agent's past interactions influence its current decision. They fit a discrete Mori-Zwanzig kernel — that's a fancy way of saying they measured how much the past matters, and for how long.

Meng: And the finding was that the current interaction weight is lower than the memoryless rate. So if an agent has a history, it's more cautious about moving. They call it conviction braking.

Lu: Exactly. And the past-interaction tail is real — it's about forty-seven percent of the current weight — but it decays quickly and vanishes after a few lags. That means you can truncate the memory and still capture the dynamics. That's huge for building efficient simulations.

Meng: So instead of tracking every agent's full history, you just need the last few steps. That's a massive memory savings when you're scaling to millions of agents.

Jane: And Meng, you're the engineer here. What does this mean practically for someone who wants to build one of these surrogate societies?

Meng: Well, the paper gives a clear recipe. You classify your simulation's perception cell — global, community, or local feed. You screen your observable to make sure it's actually measuring behavior and not an accounting identity. You elicit a few hundred to a few thousand AI decisions, fit your surrogate, and then you can run any population size on a laptop.

Tom: And they even have a scaling law for how many decisions you need. It's about a thousand to two thousand, which costs under a dollar on DeepSeek.

Meng: Right, and there's a capacity floor. If your surrogate has fewer than four features, no amount of data will fix it. But once you have enough features, more data helps. So the rule is: buy structure first, data second.

Lu: That's a really practical insight. And it connects to the cross-domain check they did with the differentiable epidemic model. Their closure-based approach recovered the planted parameters within thirteen percent and six percent, in zero point three four seconds on a laptop, versus about four hundred seconds of GPU time for the autodiff model.

Jane: That's a thousand-fold speedup for the aggregate observable. That's not a marginal improvement, that's a different regime.

Lu: It is. And it suggests that for many macroscopic questions, we don't need the full machinery of differentiable simulation. We just need the right closure.

Meng: But I want to push back a little on the TwinMarket result. They tried to reproduce financial stylized facts — fat tails, volatility clustering — and the surrogate trader alone couldn't do it. They needed the market mechanism, specifically the price-impact coupling, to be strong enough.

Jane: So that's a case where the agent isn't the bottleneck, the environment is.

Meng: Exactly. And they report it honestly as a limitation. You can't always just elicit the agent and expect the macro behavior to emerge. Sometimes you need the right market structure too.

Tom: That's a good caveat. So the improvements here are both methodological — better closures, better memory handling — and practical — clear recipes for when to use what.

Jane: And I think that's the real value. This paper isn't just a single result. It's a toolkit. And that's what we should be talking about in the conclusion.

Conclusion: Tom: So we've spent the whole show on "Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop." Let me try to pull it all together.

Jane: The core idea is that you can replace expensive AI agents with cheap mathematical surrogates, fitted from a few hundred real AI decisions, and still reproduce macroscopic behavior. But only when the perception structure allows it.

Tom: And the paper gives you a taxonomy to predict when it works. Global feeds, mean-field, error shrinks with population. Community feeds, block structure, error floors. Local feeds, graph structure, error grows.

Jane: And the two blind tests — one on contact graphs, one on real LLM responses — mostly confirmed the predictions. The two refuted predictions were traced to response curvature, and the theory matched quantitatively with no free parameters.

Lu: I think the deepest contribution is that the surrogate becomes a measurement instrument. The fitted parameter that reproduces the Phillips curve tells you what causes it. The reasoning ablation tells you that thinking, not wording, creates the macro law.

Meng: And from an engineering standpoint, the recipe is clear. Classify, screen, elicit, fit, scale. A few dollars of API calls, a laptop, and you can study million-agent societies.

Jane: There are honest limitations. Some targets need the market mechanism, not just the agent. Some observables are accounting identities. And the exact macroscopic value is a model-and-prompt fingerprint, not a universal constant.

Tom: But the bigger message is that we don't need to spend a fortune to learn something real. The field of AI-agent simulation has been gated by cost. This paper opens it up.

Lu: And that has cultural implications too. When simulation becomes cheap, more people can ask questions about how societies behave — not just big labs. That's democratizing science.

Tom: Well said, Lu. So we're going to say goodbye to this paper. It was a pleasure. The taxonomy, the blind tests, the measured memory kernel — it's a lot of substance packed into a humble title.

Jane: And the title really does say it all. Poor man's agentic modeling. You don't need a rich lab to do rich science. You just need a good idea and a laptop.

Tom: Thanks for listening, everyone. Next up, we've got a paper on something completely different, so stay tuned. Goodbye, "Poor Man's Agentic Modeling." You were a great guest.