CD Consulting R&D

KB-ECON · Entry 1 · Published 10 September 2026 · Version 1

Agent-based Computational Economics: markets made of software agents, from the Iowa State laboratories of 2001-2005 to central banks and language-model agents in 2026

Twenty years ago Leigh Tesfatsion's school at Iowa State asked what a market economy looks like once the auctioneer is removed and every firm, household and trader is a program that learns. This entry restates the founding answer and records where it has led: into the stress tests of central banks and, since 2023, into simulations where each agent is a language model.

If you model an economy with one representative agent, here is what changes when you give it thousands — from 2001 to today's LLM agents.

Synthesis of nine academic papers and lecture notes from 2001-2005, read in full and restated in our own words, then updated with public web sources consulted on 10 September 2026. The founding corpus and the recent literature are always distinguished in the text; figures are quoted with their source and nothing is estimated.

Contents

  1. 1. Why remove the auctioneer
  2. 2. What an ACE model is
  3. 3. Two laboratories
  4. 4. How little intelligence does a market need?
  5. 5. The limits the founders declared
  6. 6. Where the field stands in 2026
  7. 7. What to take away
  8. 8. Sources

1. Why remove the auctioneer

The founding texts start from a complaint about the standard model of a market economy. In the Walrasian tradition, firms and consumers take prices as given, optimise in isolation, and a fictitious auctioneer finds the prices at which every market clears before any trade happens. Tesfatsion's 2005 chapter for the Handbook of Computational Economics describes this device as a coordination mechanism that eliminates the very possibility of strategic behaviour: agents are linked through prices and through nothing else [S2]. Her lecture notes of 2003 make the same point with an old joke about a physicist, an engineer and an economist stranded with a can of beans, the economist proposing to assume a can opener; the can opener of textbook price theory is the market-clearing price itself [S8].

The objection is not that the equilibrium model is wrong in its own terms. It is that it does not say how production, pricing and trade actually happen, and that it sits badly with the demand, voiced by Kenneth Arrow as early as 1959, that prices should emerge from what the agents inside the model do [S8]. Remove the auctioneer and the modeller must immediately confront asymmetric information, search costs, strategic pricing, learning under uncertainty, rationing when markets do not clear, bankruptcy, and the possibility that the economy settles into an equilibrium that everyone would prefer to leave [S2] [S8]. Tesfatsion's game-theory glossary of 2003 gives that last outcome its precise name: a coordination failure is a Nash equilibrium that is Pareto-dominated, a state where no one gains by moving alone although all would gain by moving together [S5].

2. What an ACE model is

The corpus defines Agent-based Computational Economics, in words repeated across the papers, as the computational study of economies modelled as evolving systems of autonomous interacting agents [S1] [S2]. Four ideas give the definition its content.

First, an agent is a bundle of data and methods, in the sense of object-oriented programming, that may stand for an individual, a firm, a market, a regulator, a crop or the weather; agents can contain other agents, so a firm may be made of workers and managers [S2]. Second, the model must be dynamically complete: once initial conditions are set, the world must be able to run on the interactions of its agents alone, with no further intervention from the modeller. Tesfatsion calls this the culture-dish approach, the modeller preparing the dish and then stepping back to observe [S2] [S3]. Third, agents learn: they hold private, hidden processes, they update their behaviour from experience, and some of them change the way they learn. The 2005 slides put the consequence in one line: the main uncertainty facing an agent is not knowing what the other agents will do [S3]. Fourth, the object of study is the whole set of possible trajectories rather than a single equilibrium point, what the chapter calls the system's phase portrait [S2].

The chapter then sorts ACE research into four objectives that remain a useful map twenty years on: empirical understanding (can a global regularity be grown from the bottom up), normative understanding (does a proposed market design survive agents who try to game it, or, in Tesfatsion's image, fill the bucket with water and see if it leaks), qualitative insight into the self-organising capacity of decentralised markets, and methodological advancement, meaning the tools, experimental protocols and reporting standards that make such experiments reproducible [S2].

3. Two laboratories

Two working models anchor the corpus.

The Trade Network Game Laboratory, published in 2001 in the IEEE Transactions on Evolutionary Computation, is a Windows application in which buyers, sellers and dealers repeatedly search for trade partners, play a prisoner's dilemma with each partner, update their expectations and, at the end of each generation, evolve their strategies through a genetic algorithm [S1]. Its labour-market experiment pits twelve workers against twelve employers under three settings of relative job capacity, twenty runs each. The results are not bell-shaped: for every setting the outcomes fall into two or three sharply separated clusters, and under tight job capacity a quarter of the runs end in complete coordination failure, every agent having evolved into a defector and every worker persistently unemployed [S1]. The authors offer this as an answer to a live puzzle of labour economics, why observationally identical workers and firms end up with very different earnings histories: the model produces that heterogeneity from path dependence alone, with no difference in initial attributes [S1].

The ACE Trading World, the worked example of the 2005 chapter, is a two-sector economy of hash and beans in which firms post a production level and a price at the start of each period, consumers search for the lowest posted prices, firms ration excess demand through a random queue, and, decisively, firms that cannot cover their fixed costs become insolvent while consumers who cannot meet their subsistence needs die [S2]. Firms learn by a variant of Roth-Erev reinforcement learning, choosing among supply offers with probabilities that shift towards what has been profitable. The chapter draws six lessons from building it: the difficulty of constructing agents that merely survive, the primacy of survival over every other objective, the strategic rivalry that arises the moment firms set their own prices, the trade-off between exploiting current information and exploring for better offers, the scaffolding of conventions (insolvency rules, rationing rules, dividend policies) that any decentralised economy needs, and the modelling conundrum that in such a world everything seems to depend on everything else [S2]. A short accompanying set of lecture notes supplies the market taxonomy the model relies on: brokers who match without holding inventory, dealers who make the market from their own stock, and four organisations of exchange, from bilateral trade to organised exchanges combining auction and dealer features [S7].

4. How little intelligence does a market need?

A second lineage runs through the corpus, less philosophical and more experimental. In 1993 Dhananjay Gode and Shyam Sunder had shown that a continuous double auction populated by zero-intelligence traders, programs that bid at random subject only to a budget constraint, extracted almost all of the available surplus. Tesfatsion reads the result as a warning that good market performance should not be credited to the rationality of traders: it can come from the institution [S2]. Her 2004 notes on market basics define the yardsticks that this literature uses, market efficiency as the share of the maximum surplus actually extracted, and market advantage as the surplus a trader captures above its competitive share [S6].

Two papers in the corpus push the lineage forward. At Hewlett-Packard's Bristol laboratory in 2002, Walia, Byde and Cliff asked whether the auction rules themselves could be designed by evolution. They parameterised the double auction by a single number, the probability that the next quote comes from a seller, ran an evolutionary strategy over that parameter with zero-intelligence traders in the market, and mapped 441 test markets. The optimum was often a hybrid that no real exchange uses: near one where demand is steep, near zero where supply is steep, and in between along the diagonal. Their reading is that the parameter decides which side of the market plays price-maker and which price-taker [S4]. At the University of East Anglia in 2004, Bagnall and Toft tested two adaptive bidding algorithms, the memory-free Zero Intelligence Plus and the memory-based Gjerstad-Dickhaut, in sealed-bid auctions where the optimal strategy is known. Over a hundred runs of ten thousand auctions, the memory-free agent learned a margin close to but measurably short of the optimum, while the memory-based agent reached it in first-price auctions and overshot in second-price ones [S9]. The method matters as much as the result: benchmark an artificial agent against a known optimum before trusting it in a market where nothing is known.

5. The limits the founders declared

The 2005 chapter is candid about what the approach could not yet do, and the list is worth recording because it sets the agenda for the next twenty years. An ACE model needs a complete initial specification, and when feedbacks are strong, small changes to that specification can change the kind of outcome that emerges, so robust prediction demands intensive experimentation across plausible settings. It was not clear, Tesfatsion wrote, how well such models would scale to many thousands of agents. Validation against data was hard, because an experiment yields a distribution of outcomes, often multi-peaked, whereas the real world offers one time series from a poorly understood process. And in a multi-agent world there may be no best way to learn, since the value of any learning rule depends on what everyone else is doing [S2]. The chapter closes with a plea that every graduate programme in economics teach a programming language [S2]. The lecture notes of 2003 add a sober empirical reminder: in experiments with human subjects, double auctions in simple markets do converge to the competitive price, so the textbook theory keeps predictive content even where its behavioural assumptions are false [S8].

6. Where the field stands in 2026

Everything above comes from the 2001-2005 corpus. What follows comes from public sources consulted on 10 September 2026.

6.1 A field with institutions

The handbook Tesfatsion was drafting appeared in May 2006 as volume 2 of the Handbook of Computational Economics, edited with Kenneth Judd, with sixteen review chapters and six chapters of perspectives by, among others, Brian Arthur, Robert Axelrod, Joshua Epstein and Thomas Schelling [S10]. The ACE website that the 2001 paper already pointed to is still maintained from Iowa State, last updated in July 2026, and now states the approach as seven modelling principles of what its author calls completely agent-based modelling: agents defined as software entities, broad in scope, locally constructive, autonomous, with a world state that is nothing but the ensemble of agent states, a history driven only by agent interactions, and a modeller confined to setting initial conditions and observing without perturbation [S11]. Tesfatsion is today Professor Emerita, her stated research being electric power market design and the ACE platforms used to test such designs [S12].

The zero-intelligence lineage, for its part, crossed into empirical finance: in 2005 Farmer, Patelli and Zovko fitted a model of randomly placed orders to the London Stock Exchange and reported that, with a single free parameter, it explained 96 percent of the variance of the bid-ask spread across stocks and 76 percent of the variance of price diffusion [S17].

The wider community meets at the annual Computing in Economics and Finance conference of the Society for Computational Economics, whose thirty-second edition took place in Venice from 29 June to 1 July 2026 with agent-based modelling, machine learning and the economics of artificial intelligence among its topic areas [S18]. The most authoritative recent survey is the ninety-page review by Robert Axtell and J. Doyne Farmer in the March 2025 Journal of Economic Literature, which credits agent-based modelling with advances in markets, industrial organisation, labour, macroeconomics, development and environmental economics, and in finance with the understanding of clustered volatility, market impact, systemic risk and housing markets [S13]. Farmer's 2024 book for a general readership, Making Sense of Chaos, carries the same programme to a wider audience [S25]. A fifty-page review in the Journal of Simulation in February 2026 covers the application side for economic and financial markets [S19].

6.2 Inside central banks

The most concrete change since 2005 is institutional. A Bank of England staff working paper of February 2025, written by six economists from the central banks of Hungary, Spain, Italy, England and Poland, surveys the use of agent-based models across some two dozen central banks and related bodies. Its thesis is that these models became complementary tools after the global financial crisis of 2007-09 widened the remit of central banks to include macroprudential regulation, system-wide stress testing and, more recently, climate risk and digital currencies [S14]. Payment-system simulation came first; financial-stability models dominated after the crisis; work on price stability and climate appeared from 2016 and on central bank digital currencies from 2021 [S14].

Three strands stand out. In housing, the model built in 2016 by Bank of England and Oxford researchers represented renters, first-time buyers, movers and buy-to-let investors interacting through a double-auction market, and was used to examine loan-to-value and loan-to-income limits [S16]; it has since been adapted to Denmark, Italy and Spain, and Hungary's central bank now runs a one-to-one model of its housing market with four million households [S14]. In forecasting, Poledna, Miess, Hommes and Rabitsch published in 2023 the first agent-based model able to compete with vector-autoregression and dynamic stochastic general-equilibrium benchmarks in out-of-sample forecasts of macro variables, a model of Austria populated with millions of heterogeneous agents drawn from national accounts and census data, and applied to the effects of pandemic lockdowns [S15]. The Bank of Canada recalibrated that model to its own economy and, according to the 2025 survey, became the first central bank to acknowledge agent-based models formally within its modelling strategy, as specialty models for structures too complex for the core model [S14]. In financial stability, network and contagion models built by the European Central Bank, the Bank of England and others have entered system-wide stress tests, and Hungary has published results from such a model in its Financial Stability Report since 2016 [S14].

The 2025 survey also records how the field is answering Tesfatsion's 2005 limits. On scale, it reports models moving towards one-to-one representations, sometimes described as digital twins, and an analysis concluding that the optimal model size is often larger than previously thought. On calibration, it points to open-source software packages using machine-learning surrogates and genetic algorithms, and to experiments with reinforcement learning to speed the search. On behaviour, it anticipates language models being used as models of human decision-making inside agent-based models, while warning that defining behavioural rules, calibrating parameters and tracing emergent outcomes back to rules remain demanding [S14].

6.3 Language-model agents and the return of the validation question

The zero-intelligence lineage asked how little intelligence a market needs. The newest lineage asks the opposite question: what happens when each agent is a large language model. Three steps mark the road. In January 2023 John Horton proposed treating language models as homo silicus, a simulated economic subject that can be given endowments, information and preferences and then placed in experiments; the paper, revised in February 2026, reports that replications of classic behavioural experiments give qualitatively similar results to the originals [S20]. In 2023-24 a group at Tsinghua University presented EconAgent, a macroeconomic simulation in which language-model agents with profiles and a memory module decide how much to work and consume, and argued that the resulting aggregates were more plausible than those of rule-based or learning-based agents [S21]. In May 2025 del Rio-Chanona, Pangallo and Hommes ran language-model agents through laboratory market experiments with genuine feedback, each agent's decision moving the price that the others see next period. With a short memory and enough variability the agents reproduced the broad patterns seen with human subjects, including the difference between positive- and negative-feedback markets, but they displayed less heterogeneity than humans [S22].

That last finding is the hinge. A critical review in Artificial Intelligence Review in November 2025 argues that language models revive rather than resolve the old problem of validating agent-based models: they are black boxes with cultural biases and stochastic outputs, and most studies rest on face validity or on outcome measures only loosely tied to mechanisms [S23]. A March 2026 comparison of off-the-shelf models with human responses in a controlled experiment found that the models reproduce directional effects but not effect sizes, and that results vary across models [S24]. Read against the corpus, the pattern is familiar. Bagnall and Toft benchmarked their agents against a known optimum before trusting them [S9]; Tesfatsion warned that an experiment produces a distribution while the world produces one path [S2]. The language-model literature is now rediscovering both disciplines.

7. What to take away

For a reader who models risk, finance or policy, the corpus and its update leave four practical lessons.

8. Sources

The nine papers of the founding corpus (S1 to S9) were read in full from the author's own copies; they are cited by their bibliographic identity and are not reproduced here. Web sources were consulted on 10 September 2026.

  1. S1 Research (journal article). David McFadzean, Deron Stewart and Leigh Tesfatsion, "A Computational Laboratory for Evolutionary Trade Networks", IEEE Transactions on Evolutionary Computation, 2001. Read 10 September 2026.
  2. S2 Research (handbook chapter, working version). Leigh Tesfatsion, "Agent-Based Computational Economics: A Constructive Approach to Economic Theory", version dated 29 May 2005, forthcoming as chapter 1 of the Handbook of Computational Economics, volume 2, North-Holland. Read 10 September 2026.
  3. S3 Research (lecture slides). Leigh Tesfatsion, "Agent-Based Computational Economics: A Constructive Approach to Economic Theory", presentation, Iowa State University, spring 2005. Read 10 September 2026.
  4. S4 Research (technical report). Vibhu Walia, Andrew Byde and Dave Cliff, "Evolving Market Design in Zero-Intelligence Trader Markets", HP Laboratories Bristol, HPL-2002-290, 30 October 2002. Read 10 September 2026.
  5. S5 Research (lecture notes). Leigh Tesfatsion, "Game Theory: Basic Concepts and Terminology", Economics 308, Iowa State University, spring 2003. Read 10 September 2026.
  6. S6 Research (lecture notes). Leigh Tesfatsion, "Market Basics for Price-Setting Agents", Iowa State University, 7 February 2004. Read 10 September 2026.
  7. S7 Research (lecture notes). Leigh Tesfatsion, "Market Organization with Price-Setting Agents", Economics 308, Iowa State University, 7 February 2004. Read 10 September 2026.
  8. S8 Research (lecture notes). Leigh Tesfatsion, "Notes on Price Discovery with Price-Setting Agents", Iowa State University, 3 February 2003. Read 10 September 2026.
  9. S9 Research (conference paper). A. J. Bagnall and I. E. Toft, "Zero Intelligence Plus and Gjerstad-Dickhaut Agents for Sealed Bid Auctions", School of Computing Sciences, University of East Anglia, 2004. Read 10 September 2026.
  10. S10 Primary (publisher's table of contents, mirrored by the editor). Leigh Tesfatsion and Kenneth L. Judd (eds), Handbook of Computational Economics, Volume 2: Agent-Based Computational Economics, Elsevier/North-Holland, May 2006. https://faculty.sites.iastate.edu/tesfatsi/archive/tesfatsi/hbace.htm — consulted 10 September 2026.
  11. S11 Primary (author's website). Leigh Tesfatsion, "ACE: A Completely Agent-Based Modeling Approach", Iowa State University, page last updated 17 July 2026. https://faculty.sites.iastate.edu/tesfatsi/archive/tesfatsi/ace.htm — consulted 10 September 2026.
  12. S12 Primary (faculty page). Leigh S. Tesfatsion, Professor Emerita of Economics, Iowa State University. https://faculty.sites.iastate.edu/tesfatsi/ — consulted 10 September 2026.
  13. S13 Research (journal article, abstract page). Robert L. Axtell and J. Doyne Farmer, "Agent-Based Modeling in Economics and Finance: Past, Present, and Future", Journal of Economic Literature 63(1), March 2025, pp. 197-287. https://www.aeaweb.org/articles?id=10.1257/jel.20221319 — consulted 10 September 2026.
  14. S14 Official (central bank staff working paper). András Borsos, Adrian Carro, Aldo Glielmo, Marc Hinterschweiger, Jagoda Kaszowska-Mojsa and Arzu Uluc, "Agent-based modeling at central banks: recent developments and new challenges", Bank of England Staff Working Paper No. 1,122, February 2025. https://www.bankofengland.co.uk/working-paper/2025/agent-based-modeling-at-central-banks-recent-developments-and-new-challenges — consulted 10 September 2026.
  15. S15 Research (journal article, abstract page). Sebastian Poledna, Michael Gregor Miess, Cars Hommes and Katrin Rabitsch, "Economic forecasting with an agent-based model", European Economic Review 151, 2023, article 104306. https://econpapers.repec.org/article/eeeeecrev/v_3a151_3ay_3a2023_3ai_3ac_3as0014292122001891.htm — consulted 10 September 2026.
  16. S16 Official (central bank staff working paper, described on the co-authors' institute page). Rafa Baptista, J. Doyne Farmer, Marc Hinterschweiger, Katie Low, Daniel Tang and Arzu Uluc, "Macroprudential policy in an agent-based model of the UK housing market", Bank of England Staff Working Paper No. 619, 2016. https://www.inet.ox.ac.uk/news/bank-of-england-wp — consulted 10 September 2026.
  17. S17 Research (journal article, abstract record). J. Doyne Farmer, Paolo Patelli and Ilija I. Zovko, "The predictive power of zero intelligence in financial markets", Proceedings of the National Academy of Sciences 102(6), 2005, pp. 2254-2259. https://pubmed.ncbi.nlm.nih.gov/15687505/ — consulted 10 September 2026.
  18. S18 Official (learned society). Society for Computational Economics, "32nd CEF Conference", Ca' Foscari University of Venice, 29 June to 1 July 2026. https://comp-econ.com/32nd-cef-conference/ — consulted 10 September 2026.
  19. S19 Research (journal article, metadata record). Ruhollah Jamali and Sanja Lazarova-Molnar, "Agent-based modeling and simulation for economic markets: a comprehensive review of applications, challenges, and opportunities", Journal of Simulation, published online 19 February 2026. https://doi.org/10.1080/17477778.2026.2625187 — consulted 10 September 2026.
  20. S20 Research (preprint). John J. Horton, Apostolos Filippas and Benjamin S. Manning, "Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?", arXiv 2301.07543, first version January 2023, revised 26 February 2026. https://arxiv.org/abs/2301.07543 — consulted 10 September 2026.
  21. S21 Research (conference paper, preprint). Nian Li, Chen Gao, Mingyu Li, Yong Li and Qingmin Liao, "EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities", ACL 2024, arXiv 2310.10436, version of 24 May 2024. https://arxiv.org/abs/2310.10436 — consulted 10 September 2026.
  22. S22 Research (preprint). R. Maria del Rio-Chanona, Marco Pangallo and Cars Hommes, "Can Generative AI agents behave like humans? Evidence from laboratory market experiments", arXiv 2505.07457, 12 May 2025. https://arxiv.org/abs/2505.07457 — consulted 10 September 2026.
  23. S23 Research (journal article, abstract and open-access record). Maik Larooij and Petter Törnberg, "Validation is the central challenge for generative social simulation: a critical review of LLMs in agent-based modeling", Artificial Intelligence Review, 18 November 2025. https://doi.org/10.1007/s10462-025-11412-6 — consulted 10 September 2026.
  24. S24 Research (preprint). Adnan Hoq and Tim Weninger, "Evaluating LLMs as Human Surrogates in Controlled Experiments", arXiv 2604.15329, March 2026. https://arxiv.org/abs/2604.15329 — consulted 10 September 2026.
  25. S25 Press (publisher's page). J. Doyne Farmer, Making Sense of Chaos: A Better Economics for a Better World, Allen Lane / Yale University Press, 2024. https://yalebooks.yale.edu/book/9780300283327/making-sense-of-chaos/ — consulted 10 September 2026.

Limits. The founding corpus is one school's view of its own field: eight of the nine documents are by or with Leigh Tesfatsion, and the ninth family (zero-intelligence traders) is represented by two papers only; other agent-based traditions of the period, notably the Santa Fe artificial stock market and the European evolutionary-economics work, are seen here only through her citations. The 2026 update rests on abstracts, metadata records and one full working paper, not on a systematic literature search; the survey of central banks [S14] is itself written by central bank staff and may overstate adoption. Preprints [S20] [S22] [S24] have not all passed peer review, and the language-model results they report depend on the particular models tested and may not hold for later ones. No figure in this entry was computed by us; each is quoted from its source, and the reader is invited to check the originals before relying on any of them.