The Future of Agent-Based Modeling

The Evolution of LLM Agents

Module 11 introduced the Stanford “Generative Agents” (2023) — language-model characters in a simulated town whose answers to interview questions human evaluators found, in Stanford HAI’s summary, “more believable than responses given by humans who were pretending to be the agents” — and the validation problem such agents raise. Here we look forward.

A module of forecasts owes the reader its terms up front. Reliable long-range prediction in complex systems is highly unlikely — agents adapt, and each adaptation devalues part of the record a forecast was fitted to. The quip that predictions are difficult, especially about the future — usually credited to Niels Bohr, though it began as an anonymous Danish saying — is a working constraint here: this module maps where agent-based modeling itself is headed before Module 16 assembles the synthesis, and every “by 2028” below is a trajectory that can bend, not a schedule.

In late 2024, follow-up work by the same Stanford group showed that LLM Large language model (LLM) An AI model trained on vast amounts of text to predict the next word-piece. Chatbots such as ChatGPT and Claude are built on them. agents built from in-depth interview transcripts with more than 1,000 Americans could replicate those people’s answers to the General Social Survey 85% as accurately as the people themselves did two weeks later — measured, that is, relative to how consistently those individuals answered the same questions themselves over time, a noisy human baseline rather than a fixed ground truth, which makes the figure less absolute than it first sounds. The emerging research landscape suggests agents now operate at three increasingly sophisticated levels of capability (not to be confused with the three tiers of validity that Module 11 set out):

This capability ladder is not static — it is an active research frontier. Work presented at NeurIPS 2024 trained a generative model of human partners and used it to train agents that cooperate better with real people, tested on the cooperative cooking game Overcooked. Domain-specific applications have expanded beyond social simulation into software engineering, analytical writing, and operational workflows.

Democratization: ABM for Everyone

The most profound impact of AI on ABM may not be smarter agents but wider access. Building a model used to mean learning a programming language — NetLogo’s own, or Python. Now practitioners can describe agent behavior in plain English and let a language model write the code. Proposed formats such as Agent-Flavored Markdown (AFM) standardize that description. An AFM file is a short header of settings (which AI model to use, which tools the agent may call) followed by ordinary written instructions: what the agent can see, what it can do, and when it acts. The “markdown” is simply the light formatting of headings and bullet points that many note-taking apps use. Because the file does not depend on any one software package, the same description can, in principle, run on different platforms.

No-code platforms have followed — Google’s Agent Designer is one of several that turn a plain-language description into a working agent. The economic barrier has fallen in parallel. Module 14 tracked a 286-fold fall in the price of a fixed level of LLM capability: what a model scoring as well as GPT-3.5 on a standard knowledge test cost per token in November 2022, against October 2024. That fall moved LLM-powered simulation from a laboratory demonstration to something a research budget can contemplate; the risk section below prices a full run.

What does mainstream adoption look like? Three scenarios are plausible by 2028: data-light domains (organizational simulation, scenario planning, creative applications) will see rapid adoption; hybrid human-AI simulation will emerge as a collaborative tool; and educational expansion will bring LLM-powered ABM into university curricula. The persistent bottleneck remains validation — ABM will not replace controlled experiments, but it will become routine for theory development and exploratory modeling.

The timeline below compresses this story into four era cards. Each card prices two different kinds of run, because they buy different things. A classical run, with rule-based agents, buys computer time. A run with LLM agents buys tokens Token The unit a language model reads and writes — roughly a word-piece. A page of text is about 500 tokens; training data is measured in trillions of them. (the word-pieces a language model reads and writes), and its bill is set by the price per token. The cost curve prices this module’s running example — 10,000 LLM agents for 100 time steps, about a billion tokens — at that fixed GPT-3.5 level of capability. In November 2022 the run would have cost about $20,000: a billion tokens at $20 per million. Before you toggle the curve, commit to a guess: two years later, did the same run cost half as much, a tenth, or less than a hundredth?

Less than a hundredth. By October 2024 the run cost about $70 — a billion tokens at $0.07 per million, a fall of 286×. The chart’s vertical axis is a logarithmic scale Logarithmic scale A chart axis on which each gridline is ten times the one below (1, 10, 100, 1,000), so a steady percentage fall draws as a straight line and a thousandfold change fits on one small chart. , where each gridline is ten times the one below; that is how a 286-fold drop fits on one small chart. Two cautions keep the number honest. First, the curve holds capability fixed at the 2022 level, and few people build agents on a 2022-class model today. On current models the same run costs $500–$5,000, as the risk section works out. The 286× is the price of standing still. Second, the curve says nothing about classical runs, which never bought tokens; their cost tracks computer hardware and has fallen steadily, without a cliff. The cards’ who column widened with both prices — first through Python, later through plain English — this section’s claim that access follows price, in miniature. A price that can fall 286× in two years can also stall, which is why every date in the Road Ahead is written in pencil. The overview in the panel below is enough for everything that follows; open the detailed view only if you want the AFM mechanics.

The Convergence of Two Worlds

Two parallel ecosystems have developed with limited interoperability. Traditional ABM frameworks — NetLogo is the best-known — are built to step many simple agents through time in a controlled order, with rich visualization and mature communities, but they have no natural place for an agent that reasons in language. Modern multi-agent AI frameworks like LangGraph are built to coordinate a handful of LLM agents that reason step by step and call outside tools, but they handle large populations and simulated environments poorly.

As of 2026 the two are converging — industry pressure favors interfaces that let an agent built in one framework compose with workflows in another. Hybrid architectures are emerging that combine classical ABM state management (environment, discrete events) with LLM agent reasoning (perception, planning, dialogue).

By 2030, the distinction between “ABM frameworks” and “AI agent frameworks” may become largely semantic. Agents will have symbolic state, perceptual interfaces, cognitive modules (rule-based, neural, or LLM-based), action interfaces, and environments — regardless of which tradition built the tool.

Digital Twins: Simulation Meets Reality

Digital twins Digital twin A continuously updated computer model of one specific real thing, such as a jet engine, a factory or a city, used to monitor it and to test changes before making them. — digital replicas of physical or social systems updated in real time — represent a natural evolution from offline simulation to operational decision support. Rather than one-time simulations with fixed parameters, ABM is increasingly embedded in feedback loops Feedback loop A chain of cause and effect that loops back on itself. A reinforcing loop amplifies change (panic selling triggers more selling); a balancing loop damps it (a thermostat). with IoT sensors and real-world data streams.

Smart cities are the primary early-adopter domain. One market-research firm projects the digital-twin market as a whole — not cities alone — growing roughly tenfold, from about $25 billion in 2025 to about $260 billion by 2032. Such CAGR Compound annual growth rate (CAGR) The steady yearly growth rate that would carry a value from its start to its end over a period. A market growing at a 25% CAGR roughly doubles every three years. estimates assume sustained investment rather than reflecting any modeled mechanism; read them as directional, and with the question every forecast deserves: who benefits from the prediction? Growth curves are, among other things, the product market-research firms sell. Applications include traffic management using 3D/4D spatial data and real-time sensor inputs, dynamic optimization of energy grid load, and real-time simulation of disease spread or disaster scenarios. New Zealand’s public-health science institute, ESR (since renamed PHF Science), grew its COVID-19 modelling into ALMA, an agent-based digital twin of the whole country — more than 5 million simulated residents, each with a daily schedule — now used to test scenarios from disease spread to climate impacts.

Supply chain digital twins use LLM agents representing suppliers, logistics providers, and manufacturers who reason about disruptions and adapt strategies. Real-time integration with inventory systems enables dynamic rerouting. Healthcare systems feed patient behavior models with EHR data and wearable sensors to simulate disease progression and treatment outcomes.

The technical architecture requires: event streaming for continuous data ingestion, fast agent reasoning (an inference budget of 100–500 milliseconds per agent decision — tighter than the 0.5–2 seconds an off-the-shelf LLM call takes, as the scaling section shows), persistent queryable agent state, and feedback loops where simulation results influence real-world decisions within operational timescales. Edge computing (running agents locally) combined with cloud orchestration (coordinating large agent fleets) is becoming standard. This represents the convergence of ABM, IoT, and the AI infrastructure explored in Module 10.

If the four-layer architecture is more plumbing than you need, skip the detailed view below — nothing downstream depends on it.

Scaling: From Thousands to Billions

GPU-accelerated ABM has matured from research curiosity to practical infrastructure. Module 11 showed the classical side: FLAMEGPU 2 runs millions of rule-based agents on one GPU workstation (the tool landscape), and a 2023 economic simulation ran 331 million agents on 128 processor cores (the road ahead). The BioDynaMo platform has since run 1.7 billion classical agents on a single server, and its distributed successor, TeraAgent, about half a trillion across many servers. For agents like these, scale is a matter of hardware, and hardware keeps getting cheaper per calculation: computation per dollar on GPUs has doubled roughly every two and a half years.

LLM agents hit a different bottleneck. Take this module’s running example: 10,000 agents simulated for 100 time steps. A single off-the-shelf LLM inference takes 0.5–2 seconds; at 1 second each, one time step of 10,000 decisions made one after another takes 166 minutes. Sending the calls in parallel shortens the wait but not the bill: every one of the million decisions is still paid for in tokens, as the risk section prices out. For LLM-powered ABM the binding constraints are latency and token cost, not raw compute.

Practical solutions are emerging. Mixture of agents combines expensive LLM agents (10–100) with cheap rule-based agents (1M+) to model human heterogeneity while preserving scale. Agent-of-agents uses a single LLM agent to represent a population segment, reducing the agent count while maintaining behavioral diversity. Cached embeddings store each agent’s memories in a pre-processed numerical form, so they need not be re-read from scratch at every step. Smaller specialized models (7B–13B parameters) replace 70B+ models for faster inference at acceptable quality.

Risks, Challenges, and the Validation Frontier

Module 11 established that validation is the central challenge for LLM-powered ABM. Here we deepen the analysis. A 2025 critical review of 35 generative-agent studies in Artificial Intelligence Review found that nearly all rely on zero-shot Zero-shot prompting Asking an AI model to do a task with no worked examples in the prompt, relying on the instruction alone. prompting without fine-tuning Fine-tuning Adjusting an already-trained model on a smaller, targeted dataset so it performs a specific task or follows instructions better. , and that most validation stops at face validity (does the behavior look right?) — 22 of the 35 used subjective believability as their main check — rather than climbing to behavioral validity (do decisions and patterns match measurements?) or predictive validity (does the model forecast what it was not fitted to?) — the three validity tiers Validity tiers (face, behavioral, predictive) Three increasingly strict tests of a simulation. Face validity: does the output look right to an expert? Behavioral validity: do the agents’ decisions and the patterns they add up to match measurements? Predictive validity: does the model forecast what it was not fitted to? that Module 11 laid out. Most report results from a single simulation run. A separate study found that a language model could predict the sign of effects in its own simulated experiments (price increases reduce demand) but not their magnitude — limiting such agents’ value for quantitative policy analysis.

Bias and fairness compound the challenge. Module 11 noted that LLM agents inherit the biases of their training data; reviews catalogue the forms — agents associating preferentially with similar others, stereotyped decisions, deference to wealth and status, and thin representation of non-Western perspectives. If ABM is used for policy exploration — simulating the impact of new policies on vulnerable populations — biased agents could lead to systematically wrong conclusions about equitable outcomes.

Computational costs and environmental impact are non-trivial for LLM agents. Price the running example: 10,000 LLM agents over 100 time steps is one million agent decisions, and at roughly a thousand tokens of context, memory, and response per decision, about a billion tokens — which at 2026 input-token list prices of $0.50 per million for an efficient model (Google’s Gemini 3 Flash) to $5 for a frontier one (Anthropic’s Claude Opus 4.6 or 4.7) comes to $500–$5,000 per run — more once the pricier output tokens are counted. Every step of that multiplication is checkable by hand. A rule-based version of the same model buys no tokens at all; its bill is machine time, which is why the Builder below prices rule-based agents far below LLM-powered ones. Large-scale simulation with millions of LLM agents pushes costs further.

Where the LLM bill goes next depends on which price you follow. Model distillation, cached reasoning, mixture-of-experts Mixture of experts (MoE) A design that routes each input to only a fraction of a model’s parameters (its "experts"), so a very large model can run at a fraction of the computing cost of using all of it at once. architectures, and open-source efficient implementations keep cutting the price of a given level of capability. An illustrative estimate, not a forecast: suppose that price falls tenfold a year, slower than the roughly seventeenfold a year implied by Module 14’s 286× in two years. Then today’s $500–$5,000 run, on today’s models, would cost $5–$50 by 2028. A modeler who always upgrades to the newest frontier model gets no such guarantee, because the 286× was measured at fixed capability, not at the frontier.

Overfitting presents a subtler risk. LLM agents trained on historical behavioral data may not generalize to novel scenarios — an agent calibrated on COVID-era mobility patterns may fail for a different epidemic with different transmission routes. This is the extrapolation trap wearing a modeler’s clothes: in 1966, Harry Harrison projected his era’s trend lines thirty-three years forward and got a 1999 New York of 35 million starving inhabitants — the real figure was 8 million, because the contraceptive pill and the price mechanism were already bending his lines as he wrote (Module 14 walks the failure in full). An agent calibrated on yesterday’s data extends yesterday’s lines at machine speed. Prompt-fitting (optimizing prompts rather than model weights) is harder to regularize than traditional overfitting.

Two notes before you build. First, the Builder’s three validation buttons are Module 11’s three tiers, and its Validation score measures confidence achieved, not ambition: selecting Predictive Validity — the gold standard — lowers the score, because the stricter the standard, the less of it any configuration can reach. Second, a prediction to commit to: the City Digital Twin preset runs 10 million reinforcement-learning agents at a tenth of a cent per agent-step, for 100 steps, on a $50,000 budget. Does the budget cover the run? Multiply it out, then click. (The Builder prices an LLM agent decision at half a cent — the frontier end of the range priced above.)

If you did the multiplication, the red banner held no surprise: $1,000,000, twenty times the budget. The preset is over budget by design — city-scale ambition runs into exactly this wall. Now click Pandemic Simulator and switch its agent type from Rule-Based to LLM-Powered: Realism jumps from 30 to 90, but Interpretability falls from 90 to 20, Validation from 60 to 24, and the cost balloons from $1,000 to $500,000. No corner of the radar chart is reachable from every other corner — this section’s argument rendered as geometry. The panel’s overview carries the rest; the detailed view takes each of the three validity tiers further — how studies measure them, and how far LLM agents get.

The Road Ahead: 2026–2030

Three research frontiers — each one already under your thumb in this module’s interactives — will do the most to shape ABM’s trajectory. Validation frameworks: developing operationalizable criteria, benchmark datasets, and error bounds for generative agent simulations. Hybrid cognitive architectures: combining LLMs for planning with symbolic systems for consistency — the middle path the Builder’s agent-type trade-off points toward. Real-time digital twin integration: demonstrating operational value of live feedback from ABM to decision-makers, with interfaces for non-experts.

The sector impact predictions map onto two adoption horizons. By 2028: urban planning, public health, organizational simulation, and policy exploration — domains where ABM’s bottom-up reasoning adds immediate value and validation requirements are manageable. By 2029: financial regulation, supply chain logistics, and environmental modeling — domains with established ABM traditions where LLM agents add behavioral realism.

Will traditional tools survive? Yes, but repositioned. NetLogo’s strength in education is unlikely to be disrupted — it remains the easiest entry point for teaching ABM concepts. By 2030, NetLogo may reposition as an “ABM specification and visualization layer” atop cloud engines, rather than a standalone simulator. The value proposition shifts from simulation speed to ease of specification, visualization, and exploration.

This module has argued throughout that validation is the gate. The explorer below turns that claim into a prediction you can test. Its “Most Likely” preset sets technology adoption to 55, validation progress to 50, and cost reduction to 300-fold — and under those settings, one of the three domain groups goes completely dark. Which one? Commit to an answer, then look.

The high-stakes group. All four of its bars sit at zero. Drag Validation Progress from 50 toward 75 and the dark bars switch on one at a time, climate policy first and financial regulation last (its threshold is 70): adoption pressure and cost reduction multiply against validation, and cannot substitute for it. What you just watched has a name — a bifurcated landscape: exploratory and educational domains adopting LLM-ABM early for speed and narrative insight, high-stakes domains holding out until validation clears their bar. The field’s maturity by 2030 depends on three pillars — validation, interpretability, and fairness — and these determine whether ABM becomes a trusted tool for consequential decision-making or remains a powerful exploratory methodology with known limitations.

The trajectory connects forward to Module 16’s synthesis: ABM is not just a modeling technique — it is the computational expression of the complex perspective. The same insight that drives complexity science — that macro-level behavior emerges from micro-level interactions — is what makes ABM indispensable for understanding economies, societies, epidemics, and the AI systems that increasingly shape them all.