AI agents and ABMs: the science already simulating your customers

AI agents no longer just answer questions: they've started to simulate people. Stanford builds computational replicas of 1,052 real individuals that answer surveys the same way their originals do 85% of the time. Google DeepMind maintains an open library for simulating entire societies. MIT recreates, agent by agent, the 8.4 million inhabitants of New York. And brands are beginning to rehearse launches, messages and prices against synthetic populations before spending a euro in the real world. If you work in marketing, product or strategy, this lands squarely on you: your customers are already being simulated, and the next time someone sells you "synthetic consumers" you'll want to know what's behind it, what actually works and what's smoke. What's behind it has a name, thirty years of science and founding papers you can read: agent-based models, ABMs.

And I can tell this story from the inside. In the early 1990s, in the Computer Architecture unit at the Universitat Autònoma de Barcelona, exactly two of us believed in simulating the world this way, bottom-up, individual by individual: Remo Suppi and me. And by 1994 this was practice, not blackboard: we were working on ARGOS, a European project with Hatfield Polytechnic to guide an autonomous vehicle on networks of transputers, the British microprocessors of massive parallelism in which each chip was an autonomous unit with its own memory and its own message channels. The metaphor was double and we didn't know it: the car was an agent perceiving and deciding in its environment, and it was driven from the inside by a swarm of processor-agents passing messages to one another. My half of the story is still published on this very site, my 1993 postgraduate research on predicting the performance of parallel systems through simulation. The smoking gun, with a date on it.

The idea we defended had no name. Ecologists called it "individual-based models", social scientists said "microsimulation", and AI people confused it with distributed systems. When you tried to explain it, the usual response was a blend of politeness and pity: those are just programs with objects, that doesn't scale, that can't be validated, that isn't serious science. Serious science meant differential equations, aggregate models, the representative agent. We believed in something else: in populations of agents with simple rules, interacting, out of which emerged behavior nobody had programmed.

Thirty-three years later, that corridor heresy is the frontier of AI: even central banks calibrate agent models for financial stress. And at 498A, our R&D lab, we've turned the first stretch of that old conviction into a product already used by more than sixteen brands.

This article is three things at once, and all three are personal. An honest explanation of what an agent-based model is and of the 1993-96 window in which the thing gained a name, a manifesto and a toolkit, told through the original white papers. A chronicle of what happened to the other believer from that corridor, who never got off the boat. And the map of how that tradition flows, thirty years later, into what we're building with GEORadar: simulating the decision-making of synthetic people. I'll also tell you what we don't know how to do yet, because in this field anyone who doesn't tell you the limits is selling you smoke.

Two ways to explain a system

The whole story fits in one dichotomy. When you want to understand a system made of many parts, a market, a city, an epidemic, the perception of a brand, there are two roads.

The classical road goes top-down. You write equations over the aggregates: total demand, average speed, infection rate. It works wonderfully when the parts are identical and interchangeable, like molecules in a gas. And it breaks, silently, when the parts are heterogeneous, when they learn, imitate each other and react to one another. Whatever doesn't fit in the equation stops existing: the bubble, the phantom traffic jam, the rumor that goes viral.

The opposite road goes bottom-up. You define a population of autonomous agents, each with simple local rules: it perceives its environment, decides, acts, interacts. You run the simulation thousands of times and watch which macroscopic pattern emerges without anyone having explicitly programmed it. That's an agent-based model, an ABM. It isn't a point-prediction technique; it's another epistemology. You don't ask "what number will come out?", you ask "what futures can grow from these rules?".

Two epistemologies: describe the aggregate or grow the phenomenon. Source: own recreation based on Epstein & Axtell, 1996, and Schelling, 1971.

The founding example didn't even need a computer. In 1971, Thomas Schelling studied urban segregation by moving coins on a board. Each coin followed a single rule: if fewer than a third of my neighbors are like me, I move. A feather-light individual preference, almost tolerant. The result, after a few rounds, was a brutally segregated board. Nobody wanted segregation; segregation emerged anyway. That's the intellectual chill of ABM: the macro is not the sum of the micro, it's its non-obvious consequence. Then came Conway's Game of Life (1970) and Craig Reynolds' Boids (1987), perfect bird flocks generated by three local rules, with no leader and no choreographer.

By the time I arrived at UAB, all of this existed as scattered curiosity. What didn't exist was the field.

1993-1996: the window in which the thing gained a name, a manifesto and a toolkit

The transformation from curiosity into discipline happened inside a four-year window, and you can document it white paper by white paper. Most of it came out of the Santa Fe Institute, the monastery of complexity science, as working papers that are still freely available today. I've retrieved them all, and rereading them is an exercise in humility: almost everything that looks new today was already there, waiting for computing power and a decent artificial mind.

Four years to go from not existing to having a term, a manifesto, a toolkit and the first killer apps. Source: own recreation.

1993. The theory before the name. David Lane publishes Artificial Worlds and Economics as an SFI working paper: the first serious conceptual frame for "artificial worlds", bottom-up simulations to study how hierarchical organization emerges in an economy. The full text reads today like a premature manifesto.

1994. The year of the baptism. Robert Axtell and Joshua Epstein publish in the SFI Bulletin a short piece with a title that says it all: Agent-Based Modeling: Understanding Our Creations. It's the birth certificate of the term. That same year, Brian Arthur presents the thought experiment that would become the "hello world" of complexity economics: the El Farol problem, an SFI working paper that the American Economic Review published as Inductive Reasoning and Bounded Rationality.

El Farol deserves two paragraphs, because it's the direct seed of what we do today. It's a bar in Santa Fe with Irish music on Thursdays. A hundred people want to go, but it's only enjoyable if fewer than sixty show up. No communication or coordination is possible: everyone decides at home, with their own theories about how many people will go. If everyone reasoned the same way, deductively, the system would explode: everyone would go or no one would. Arthur equipped his hundred agents with inductive reasoning: each one keeps a small repertoire of predictive models, uses the one that's been working and discards it when it fails.

A hundred private theories competing produce an order that none of them contains: mean attendance locks onto the capacity. Source: own recreation based on Arthur, AER 84(2), 1994.

The result is hypnotic: mean attendance self-organizes around sixty, exactly the comfortable capacity, without ever converging to a static equilibrium and without any agent knowing the full system. A market in miniature, with genuine bounded rationality. And here comes the wink that makes me love it even more: in February 2026, a team put agents back into Arthur's bar, this time with LLMs inside, and out of the conversations emerged tribalism worthy of Lord of the Flies. Thirty-two years later, the same experiment keeps producing surprises. That's the mark of a good model.

1994 didn't end there. Palmer, Arthur, John Holland, Blake LeBaron and Paul Tayler published in Physica D the Santa Fe artificial stock market: adaptive agents buying and selling a stock, with bubbles, crashes and persistent volume as emergent properties, not as assumptions (open access text; the mature version, Asset Pricing Under Endogenous Expectations, arrived in 96). If I had to point at GEORadar's direct ancestor today, it would be this paper: replace "asset price" with "brand narrative" and "agents' expectations" with "the AI engines' perception" and the conceptual skeleton is the same. In parallel, in Europe, Nigel Gilbert and Jim Doran were founding the social school with Simulating Societies, and Kathleen Carley and Michael Prietula opened the organizational branch with Computational Organization Theory: organizations as collections of task-oriented adaptive agents.

1995. The definition, and the fertile schism. Michael Wooldridge and Nick Jennings publish the most cited survey of the era, Intelligent Agents: Theory and Practice, and fix the canonical definition of an agent: autonomy, reactivity, pro-activeness, social ability. There the field forks into two sibling branches that are worth not confusing, because the confusion is still alive today. Multi-agent systems (MAS) are engineering: agents that solve problems, negotiate, orchestrate. ABM is science: agents that explain phenomena through emergence. One builds systems; the other builds understanding. That same year, Gilbert and Rosaria Conte publish Artificial Societies, now open access, with Gilbert's chapter on emergence in social simulation, still the best conceptual introduction I know. And a home note: the MAS branch scored a milestone very early on our own campus, because between 1996 and 1998, at the IIIA-CSIC in Bellaterra, a team with Juan Antonio Rodríguez-Aguilar among its authors turned the descending-price auction of the fish market into FishMarket, an agent-mediated electronic auction house: software agents buying and selling fish inside an electronic institution.

1996. The tools and the masterpiece. Three blows in one year. Nelson Minar, Roger Burkhart, Chris Langton and Manor Askenazi publish the white paper for Swarm, the first general-purpose toolkit for building multi-agent simulations. The detail I love: among the authors are the father of artificial life (Langton, SFI) and an engineer from John Deere (Burkhart), because the tractor company already sensed this was useful for something. Until Swarm, every ABM was written from scratch, in C or Lisp, bare-handed. I know that first-hand.

Joshua Epstein and Robert Axtell publish Growing Artificial Societies (Brookings/MIT Press), the definitive manifesto. Their model, Sugarscape, is painfully elegant: a landscape of sugar, agents with vision and metabolism, rules for moving, eating, trading, reproducing. From those rules emerge migrations, wealth distributions skewed like the real ones, trade, war, cultural transmission, epidemics. The book's thesis fits in one sentence that became the motto of all "generative social science": if you didn't grow it, you didn't explain it. There's even a modern formal specification for anyone who wants to reimplement it, which is the best ABM learning exercise there is.

And Marco Dorigo, Vittorio Maniezzo and Alberto Colorni publish Ant System in IEEE Transactions: artificial ant colonies solving combinatorial optimization by depositing digital pheromone. Pure stigmergy, applied ABM founding the entire field of colony optimization.

In four years, from not existing to having a term, a definition, a manifesto, a toolkit and killer apps. And yet.

Why nobody believed: the three walls

If the 93-96 window was so fertile, why did ABM spend the next twenty-five years as a second-class citizen of science? Because it kept crashing into three walls, and all three were real. They deserve to be named with respect, because the field's recent history is exactly the history of their demolition.

The cognition wall. The agents of the 90s were toys. Four if-then rules, a state table, a genetic algorithm if you were feeling fancy. Every seminar critique was the same one, and it was fair: "your fake humans don't think like humans". The modeler had to hand-write the rules of the mind, and minds don't fit in hand-written rules. This was the deep objection, the epistemological one.

The scale wall. Simulating a few agents was trivial and simulating many was impossible. The compute of the 90s allowed hundreds, maybe thousands of simple agents. Any interesting phenomenon, a city, a real market, a pandemic, needs millions. And parallelizing a simulation where everyone interacts with everyone isn't distributing work, it's distributing conversations: the ugliest synchronization problem in distributed computing.

The calibration wall. An ABM has dozens of parameters and no derivative. Fitting it to real data was artisanal trial and error, expensive and poorly reproducible. The econometricians, with their estimable models, were right to smirk a little.

Three walls, three decades. What nobody told us in that UAB corridor is that all three would fall almost at once.

The other believer never got off the boat

Here the story turns personal, and it's the part I most wanted to write.

Of the two believers in that unit, I went through the other side of the mirror: product, brands, business, eleven years at Nike, startups, an agency. The question came with me, hibernating. Remo Suppi stayed, defended his thesis at UAB in 1996, the same year as Swarm and Sugarscape, and spent the next three decades methodically demolishing one of the three walls: scale. His entire career, seen from 2026, is a thirty-year answer to the question "what if agents could be millions?".

The published trail is beautiful to follow. In 2002, with Pere Munt and Emilio Luque, he publishes Using PDES to Simulate Individual-Oriented Models in Ecology: fish schools, the individual-oriented model par excellence, running on parallel discrete event simulation. In 2003 he adds real-time 3D animation; later fuzzy logic inside each fish's head and increasingly refined cluster-distributed versions. The fish were the perfect excuse: pure emergence, leaderless self-organization, the Boids problem elevated to serious computational science.

Then the agents became people. His group, HPC4EAS, applied all of it to emergency evacuations: panicking crowds simulated agent by agent, first as a cloud service (Crowd Evacuations SaaS, ICCS 2015), then with crowd turbulence on GPUs (ICCS 2016), and finally distilled into Care HPS, his high-performance agent simulation framework (Future Generation Computer Systems, 2017), applied among others to the evacuation of the Fira de Barcelona. The same conceptual engine ended up simulating tumor growth and urban mobility with real data and GIS maps. Fish, crowds, tumors, traffic: four skins of the same animal, the individual-oriented model, scaled with HPC.

I write this with a mix of vicarious pride and vertigo. Pride because the other believer was right and defended it with long-distance-runner constancy, paper by paper, while the field remained marginal. Vertigo because his lifelong research problem, how many agents fit in the hardware, is literally the AI headline of 2025. Now we get to that.

2023: the cognition wall falls

The unlock didn't come from where the field expected. It came from language models.

In April 2023, Joon Sung Park and colleagues from Stanford and Google publish Generative Agents: Interactive Simulacra of Human Behavior, the "Smallville" paper: twenty-five agents living in a little Sims-style town, each with an LLM as a brain and an architecture of memory, reflection and planning around it. The agents have breakfast, go to work, meet, gossip. One agent decides to throw a Valentine's Day party and, without anyone programming it, the invitations propagate, coordination emerges, someone invites their crush. Believable social behavior, emerging.

The deep reading of Smallville isn't "how cute". It's that the cognition wall, the founding objection against ABM, had just fallen. You no longer need to hand-write the rules of the agent's mind: you elicit them from a model trained on the entire written record of humanity. A synthetic persona is no longer four if-thens; it's a full profile conditioning a statistical mind. What Sugarscape did with four rules, you now do with a biography.

Serious validation arrived in November 2024, and it's the result I use whenever someone gives me the same look we got in 93. Park's team interviewed for two hours 1,052 real people, a representative sample of the US, and built one generative agent per person, fed with the interview transcript. Then they gave the agents the General Social Survey, personality tests and economic games. The result: the agents replicate their humans' answers with a normalized accuracy of 85%. The bar is not arbitrary: it's the agreement rate of the people with themselves when retaking the questionnaire two weeks later. 85% of that self-consistency. And with less bias across racial and ideological groups than agents described only with demographics.

Around that result an entire ecosystem has crystallized, and fast:

• Concordia, from Google DeepMind, is an open library for generative ABM with a tabletop-RPG "Game Master" pattern: the game master simulates the environment, the agents play (paper, experiment design guide). It's on version 2.0 and is used from social science to service evaluation.
• Project Sid, from Altera, dropped more than a thousand autonomous agents into Minecraft with their PIANO architecture: a merchant hub emerged, economic roles, a religion that spread, and a constitution the agents voted on and amended. Civilization in miniature as a tech demo.
• AgentSociety (2025) simulates more than ten thousand agents with minds equipped with emotions, needs and motivations, and logs five million interactions to study social dynamics: polarization, misinformation spread, public policy.
• GATSim brings generative agents to urban mobility, and EconSimulacra (June 2026) builds digital twins of full socio-economic systems with LLM agents.

This movement is already called GABM, generative agent-based modeling, and it's the explicit fusion of the two 1995 traditions: the mind comes from the MAS-LLM branch, the method comes from the ABM branch. The two branches of the Wooldridge-Epstein schism, finally reunified. Whenever I think about it, I end up in the same place: in the UAB corridor we were defending both at once without knowing they were two.

The other revolution: the scale and calibration walls fall

While the media spotlight went to the talkative agents, the other half of the problem, Remo's half, was being solved with less noise and the same force.

Ayush Chopra's group at the MIT Media Lab built AgentTorch, the framework for Large Population Models: agent simulations written as differentiable tensors, running on GPU. Their flagship result answers exactly the corridor question: they simulated the 8.4 million inhabitants of New York during the pandemic, agent by agent (AAMAS 2025). The trick for giving agents an LLM head without going bankrupt on API calls is elegant: archetypes. You don't call an LLM per agent; you prompt a set of representative archetypes, synthetic personas capturing behavioral segments, and their decisions govern the populations they represent (accessible explainer, Large Population Models). Hold that concept, it reappears in two paragraphs: it's exactly how a persona works in GEORadar.

And the calibration wall fell with differentiable ABM. If you rewrite the simulation so it admits derivatives, calibrating it against real data stops being craftsmanship and becomes gradient descent, like training a network. GradABM simulates populations of millions in seconds on commodity hardware and calibrates with neural networks; there's Bayesian calibration of differentiable ABMs, variational inference with Gaussian-process surrogates and a recent treatise on automatic differentiation of ABMs (September 2025). This is no longer speculative academia: central banks review in 2025 their production ABMs for systemic risk, and regional vaccination policy is calibrated with neural-network-accelerated ABMs.

The scale branch and the cognition branch, converging thirty years later. Source: own recreation.

The 2026 state of the art, said in one line: the cognition branch has achieved plausible agents and the scale branch has achieved calibratable massive populations, and the frontier is marrying them. Rich minds × millions of agents × calibration against real data. Whoever joins the three has the society simulator that in 93 was heresy.

The asterisk is missing, and in this house asterisks go in the body text, not the fine print. Validation remains open. There's serious critical literature showing that LLM-based human simulations are not yet reliable for many uses: bias toward WEIRD populations, variance flattening (agents are more homogeneous than the humans they imitate), sycophancy, identity drift over long horizons. It's being attacked from several flanks, from identity coherence with retrieval to automated causal discovery of why what emerges emerges (April 2026). It is, and this makes me smile, exactly the validation problem the econometricians threw at us in the 90s, in new clothes. The difference is that now the field has tools, money, and Stanford, DeepMind and MIT working on it. Back then there were two of us.

From measuring perception to simulating decision: the GEORadar milestone

And here the story stops being chronicle and becomes roadmap, because I'm not telling this thirty-year arc as a spectator.

Our underlying goal at 498A is the agentic simulation of human decision-making: being able to run a synthetic population against a scenario, a brand, a message, a possible future, and observe what perception forms and what decisions emerge, before spending the budget in the real world. What in my head has always been called, since that corridor, simulating futures.

GEORadar is the first milestone of that goal turned into a product, and understanding it as applied ABM explains better than anything why it works. When we audit how generative AI surfaces see a brand, we don't throw four generic questions at a chatbot. We build a synthetic demand population: personas by segment, with classified intents (informational, transactional, skeptical, comparative), traversing the funnel stages. That population generates between 3,000 and 30,000 custom prompts per study, which we run against six surfaces: ChatGPT, Gemini, Claude, Perplexity, Copilot and Google's AI Overviews. A semantic completeness algorithm detects the saturation point, the moment when asking more no longer discovers anything new, which is the GEO version of any serious simulation's convergence criterion. And what we measure at the end is exactly an emergent property: the aggregate perception, share of voice, sentiment, attributes, narrative, that no individual prompt contains but the ensemble reveals. We've run more than a million simulated prompts and analyzed over nine million brand mentions with this method.

The intellectual genealogy isn't decorative; it's operational. The concept-by-concept mapping:

And now the technical honesty, which here is part of the product just as it was in the watermarking article. What we do today is, strictly speaking, microsimulation at scale: massive independent samples. It is not yet full ABM. What defined a 96-vintage ABM was interaction with feedback: agents affecting each other and the environment, and from there strong emergence. That's exactly the next stretch of the roadmap, and every piece is a feature with a name: multi-step journeys with memory, the agent reads the engine's answer, updates its consideration set and asks again, from skepticism to comparison, like a real person; population feedback, what happens to aggregate perception when 20% of agents discard the brand at the second interaction; social dynamics, synthetic word of mouth, the between-agent influence Axelrod modeled for culture; and counterfactuals, the same population running against two different brand narratives, two measurable futures before choosing one. There, simulating futures stops being a positioning metaphor and becomes the literal description of the product.

The first stretch is already a product; the second is the roadmap. Source: own recreation, 498A.

Why tell it with this frame and not the fashionable one? Because the intellectual surname matters. GEORadar doesn't descend from SEO; it descends from Santa Fe. The practical difference between a GEO SaaS firing forty generic prompts and a study with a designed synthetic population isn't volume, it's epistemology: one runs surveys on an oracle, the other grows a phenomenon until it emerges and saturates. And because the same standards serious ABM developed over thirty years, ensembles, convergence criteria, calibration against real data, sensitivity analysis, are the yardstick we demand of any synthetic-people simulation, ours included. The day someone sells you "synthetic consumers" without talking about validation, you now know what to ask.

Coda: thirty years to become timely

Back to the 1993 corridor, because endings are written where beginnings started.

There were two of us who believed, and we were both right, but each through a branch we couldn't tell apart back then. Remo kept the hard engineering question, how many agents fit in a machine, and his answer, from fish schools on PDES to the Fira crowds on GPUs, is today the branch MIT has taken to 8.4 million synthetic New Yorkers. I took the soft question, what does an agent that resembles a person decide and what is that good for, and the answer took me thirty years, a whole career in brands and an R&D lab, but it exists, it bills, and it's called GEORadar. Today I supervise AI master's theses with UAB and IIIA-CSIC, one corridor away from where it all started, which proves that the topology of a career is also an emergent phenomenon.

What fascinates me most about the arc is how little the idea changed. Schelling with his coins in 1971, Arthur with his bar in 1994, Epstein with his sugar in 1996, Stanford with its 1,052 interviews in 2024 and us with our demand populations in 2026 are doing exactly the same thing: giving up the all-seeing god's equation and asking the anthill instead. The only thing that changed is the ant's head. Four rules then, a statistical mind trained on all of human culture now.

In 93 they told us that wasn't serious science. The answer took thirty-three years and we didn't give it: Stanford, DeepMind, MIT, the central banks and a UAB professor who never stopped publishing did. Arriving thirty years early is uncomfortable, but in innovation it's the only elegant way of arriving on time. And if this article has a practical moral, it's this: when you find an idea only two people in a corridor defend, write down the other one's name. You can recognize the good heresies because the heretics end up running the labs.

Technical sources to learn more

The founding canon, 1993-1996

• Lane: Artificial Worlds and Economics, SFI Working Paper, 1993 (full text). The theory before the name.
• Axtell, Epstein: Agent-Based Modeling: Understanding Our Creations, Bulletin of the Santa Fe Institute, 1994. The birth certificate of the term.
• Arthur: Inductive Reasoning and Bounded Rationality (El Farol), American Economic Review 84(2), 1994 (original SFI working paper).
• Palmer, Arthur, Holland, LeBaron, Tayler: Artificial Economic Life: A Simple Model of a Stockmarket, Physica D 75, 1994 (open access); Asset Pricing Under Endogenous Expectations, SFI, 1996.
• Gilbert, Doran (eds.): Simulating Societies, UCL Press, 1994.
• Carley, Prietula (eds.): Computational Organization Theory, Erlbaum, 1994.
• Wooldridge, Jennings: Intelligent Agents: Theory and Practice, Knowledge Engineering Review 10(2), 1995 (Cambridge).
• Gilbert, Conte (eds.): Artificial Societies, UCL Press, 1995, open access; Gilbert: Emergence in Social Simulation.
• Rodríguez-Aguilar, Martín, Noriega, García, Sierra: Towards a Test-bed for Trading Agents in Electronic Auction Markets, AI Communications, 1998. FishMarket, IIIA-CSIC's electronic auction house.
• Minar, Burkhart, Langton, Askenazi: The Swarm Simulation System, SFI Working Paper, 1996.
• Epstein, Axtell: Growing Artificial Societies, Brookings/MIT Press, 1996 (MIT Press; formal Sugarscape specification).
• Dorigo, Maniezzo, Colorni: Ant System: Optimization by a Colony of Cooperating Agents, IEEE Transactions on SMC-B 26(1), 1996.

The scale branch: Remo Suppi and UAB's HPC4EAS

• Suppi, Munt, Luque: Using PDES to Simulate Individual-Oriented Models in Ecology, ICCS 2002.
• Suppi, Luque: Fish Schools: PDES Simulation and Real Time 3D Animation, PPAM 2003; A Fuzzy Logic Fish School Model, ICCS 2009; High Performance Individual-Oriented Simulation.
• Crowd Evacuations SaaS (ICCS 2015), Crowd Turbulence on GPU (ICCS 2016), Care HPS (Future Generation Computer Systems, 2017): see his DBLP page and the HPC4EAS evacuations page, with the Fira de Barcelona case.
• Tashakor, Suppi: Agent-based model for tumour analysis using Python+Mesa, 2019.
• González Cuevas, Suppi: ABM simulation focused on urban mobility.
• Institutional profile: UAB Research Portal and group publications.

The generative turn, 2023-2026

• Park, O'Brien, Cai, Morris, Liang, Bernstein: Generative Agents: Interactive Simulacra of Human Behavior, 2023. Smallville.
• Park et al.: Generative Agent Simulations of 1,000 People, 2024 (Stanford HAI summary). The 85% fidelity.
• Google DeepMind: Concordia, generative social simulation library (paper; reliable experiments guide).
• Altera: Project Sid: Many-agent simulations toward AI civilization, 2024 (code).
• Piao et al.: AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents, 2025.
• GATSim: Urban Mobility Simulation with Generative Agents, 2025; EconSimulacra: A Digital Twin Platform of Socio-Economic Systems, 2026. (Recent preprints.)
• Three AI-agents walk into a bar..., 2026. El Farol revisited with LLM agents. (Recent preprint.)

Industrial scale, differentiability and calibration

• Chopra et al.: On the Limits of Agency in Agent-Based Models, AAMAS 2025 (MIT Media Lab; AgentTorch); Large Population Models, 2025.
• Chopra et al.: Differentiable Agent-based Epidemiology (GradABM), 2022.
• Quera-Bofarull et al.: Bayesian Calibration of Differentiable Agent-Based Models, 2023; Automatic Differentiation of Agent-Based Models, 2025; calibration with Stein variational inference and Gaussian-process surrogates, 2025.
• Anirudh et al.: Accurate Calibration of Agent-based Epidemiological Models with Neural Network Surrogates, PMLR.
• Agent-Based Modelling at Central Banks: Recent Developments and New Challenges, 2025; neural-network calibration of ABMs for vaccine policy, Vaccine, 2023.

Validation and critique

• LLM-based Human Simulations Have Not Yet Been Reliable, 2025. The mandatory read before believing anything.
• ID-RAG: Identity Retrieval-Augmented Generation for Long-Horizon Persona Coherence, 2025. (Recent preprint.)
• CAMO: Automated Causal Discovery from Micro Behaviors to Macro Emergence in LLM Agent Simulations, 2026. (Recent preprint.)