Agent-based modeling simulates a token economy as a population of individual actors, each with its own rules, state and reaction to what the others just did, then reads the economy off the aggregate of their decisions. It is the alternative to a closed-form model, where one demand equation stands in for everybody. The output that justifies the cost is the behaviour a spreadsheet cannot produce: a cascade, a coordination failure, a sell that starts because other people are selling.
A closed-form model tells you what happens when the average holder behaves as assumed. An agent model tells you what happens when a small minority does something the rest then react to. Token economies fail from the second case, and the first one cannot represent it.
What an agent model contains that an equation does not
Three things. Heterogeneity, because agents differ: a long-term holder, a yield farmer, a validator running on borrowed capital and a market maker do not share a utility function. State, because each agent remembers its own position, entry price and lock status. And interaction, because agents observe system-level metrics and each other, and their decisions change the metrics that the next decision reads.
That loop is the entire difference. In a closed-form model price is a function of parameters. In an agent model price is an outcome of decisions made by looking at the price, so the model can produce reinforcing behaviour nobody wrote down. Karra and co-authors set the structure out for blockchain economies: model supply dynamics from cryptoeconomy metrics, then build autonomous agents that decide on those metrics, with Filecoin as the case study.1
What it adds over a closed-form token model
Four kinds of question become answerable. Reflexive loops, where a falling staking rate raises the yield, which attracts stakers, which lowers the yield again, and whether that settles or oscillates. Threshold effects, where nothing happens until enough agents cross a decision boundary at once. Path dependence, where identical parameters give different outcomes depending on event order. And distributional outcomes, since every agent is tracked individually.
The designs that need it have feedback in them: two-sided markets where supply and demand each wait for the other, DePIN networks where operators enter on modelled yield, and game economies where a sink's effectiveness depends on how many players use it. A published agent-based study of speculative trading in token markets makes the same case, modelling investors as interacting agents rather than as an aggregate demand curve.2
Named tooling, and what each one is for
cadCAD is the framework most token engineering work is built on. It is a Python library organised around state variables, policy functions that represent agent decisions, and partial state update functions that apply them, with parameter sweeps and Monte Carlo runs built in.3 It rewards a team that already knows what it wants to model, because the structure makes you name your state before you write anything.
TokenLab is the more specific option: an agent-based token economy simulation library published alongside academic work on speculative trading, aimed at modelling and auditing token economies rather than general complex systems.2 Both rest on the same point. The tool is not the hard part. Specifying agent behaviour you can defend is.
Where it misleads, and it does
An agent model has far more free parameters than a closed-form one, and every behaviour rule is an assumption with nothing empirical behind it. You choose how a yield farmer decides to exit, and that choice largely determines the result. It is unusually easy to build a model that produces the dynamic you already believed, and unusually hard for a reader to catch it, because the mechanism is buried in agent logic rather than stated as a coefficient.
Validation is the unresolved part. With eighteen months of onchain history and one market regime there is not enough data to confirm your agent rules reproduce observed behaviour, only enough to confirm they do not contradict it. Read the output as plausible dynamics rather than probabilities, and never quote an agent model's price path as a forecast.
When we reach for it, and when we do not
We use it when the failure mode we are testing is behavioural and reflexive, when the design has two sides that bootstrap off each other, or when the question is about a minority of participants triggering a cascade. In those cases a closed-form model structurally cannot answer, so the extra parameter risk is the price of any answer at all.
We do not use it when a deterministic revenue model plus a Monte Carlo layer will do, which covers most of the work. If the question is whether emissions exceed sinks under a range of demand assumptions, that is arithmetic with uncertainty attached, and a population of simulated traders adds fragility rather than insight.
Common questions
What is agent-based modeling in tokenomics?
It simulates a token economy as a population of individual actors, each with its own rules and state, and derives system behaviour from their interacting decisions. The published blockchain version models supply dynamics from cryptoeconomy metrics, then has autonomous agents decide based on those same metrics.1 It is used when the outcome depends on how participants react to each other, which a single aggregate demand equation cannot represent.
What tools are used for agent-based token modelling?
cadCAD is the most widely used framework in token engineering, a Python library structured around state variables, policy functions and partial state update functions with parameter sweeps built in.3 TokenLab is a token-specific agent-based simulation library published alongside academic work on speculative trading in token markets.2 General-purpose agent modelling environments exist too, but the token-specific tooling saves the most setup.
When is agent-based modeling worth it over a simpler simulation?
When the behaviour you are testing is reflexive: participants reacting to each other and to metrics their own actions move. Two-sided markets, DePIN supply bootstrapping and game economies fit. If the question is whether emissions exceed sinks across a range of demand assumptions, a deterministic model with a Monte Carlo layer answers it with fewer assumptions and less room to accidentally engineer the conclusion you wanted.
See Tokenomics Data Room for how this applies in practice.
Sources
- An Agent-Based Model Framework for Utility-Based Cryptoeconomies
Kiran Karra, Tom Mellan, Maria Silva, Juan P. Madrigal-Cianci, Axel Cubero Cortes and Zixuan Zhang, arXiv:2307.15200, published in Ledger, 2023
Peer-reviewed ABM framework for blockchain economies: supply dynamics modelled from cryptoeconomy metrics with autonomous agents deciding on those metrics, with Filecoin as the case study. - Modeling Speculative Trading Patterns in Token Markets: An Agent-Based Analysis with TokenLab
Mengjue Wang and Stylianos Kampakis, University College London, arXiv:2412.07512, 2024
Introduces TokenLab, an agent-based token economy simulation library for modelling, analysing and auditing token economies. - cadCAD: Complex Adaptive Dynamics Computer-Aided Design
cadCAD project, maintained by BlockScience
The Python simulation framework most token engineering work is built on. Documentation at cadcad.readthedocs.io and source at github.com/cadCAD-org/cadCAD.
Last reviewed 2026-08
Know the terms but not sure how they apply to your project? That is what an engagement is for. We design, document, and stress-test the whole token economy inside the Tokenomics Data Room.
80+ projects advised. Complete tokenomics in 4 to 6 weeks.