AI makes prediction and calculation cheaper. It does not decide what people should value—or guarantee that privately intelligent decisions add up to a good economy.
Who should coordinate a world of AI agents when information is incomplete, incentives diverge, and capital cannot instantly move?
Follow the evidence →START WITH THE FINDINGS
The findings below are exploratory simulations—not conclusions about actual markets or real central planning.
The market–planner difference is descriptive and uncertain; the observed five-seed result does not establish a general advantage.
Inspect seed-by-seed ↗In this stylized model, withholding offset severe overproduction. Markups still increased. This runs against H2's welfare prediction.
See the adverse finding ↗The planner lost less capital than the diverse market under shared heuristic investment rules. This is a conditional result, not a planning theorem.
Explore the scenarios ↗01 / THE IDEAS
Mises, Hayek, artificial intelligence, and mechanism design illuminate distinct parts of the coordination problem. Choose a lens to see how the theories connect.
Mises begins with purposeful action. An AI agent can optimize a delegated objective, but optimizing is different from deciding which human ends deserve pursuit.
Key distinction: more computing power can reduce the cost of solving a model. It does not, by itself, create missing preferences, reveal tacit knowledge, or align incentives to report truthfully.
02 / INSTITUTIONAL DESIGN
The agents are learning, but the rules around them differ. Each institution controls who sees information, who sets allocations, and which strategic actions are possible.
03 / HOW THE EXPERIMENT WORKS
Every regime confronts the same economic environment, but receives different information and applies different decision rules.
Principals have preferences that change. Producers face resource and investment constraints.
Agents receive noisy, delayed or strategic descriptions of what people want.
Markets set prices; planners allocate; audits and payments alter incentives.
Resources move, possibly with irreversible loss and waiting periods.
The system realizes gains and losses; agents update strategies for the next period.
Percentage shortfall from a clairvoyant ex-post, perfectly informed, instantly flexible upper bound. Lower is better. It is not the welfare of a feasible real-world planner.
04 / THE EVIDENCE
Look at the recorded means, then inspect each matched random seed. These are executed synthetic runs, with small samples and exploratory status.
Compare welfare regret across five learned institutional regimes.
What these charts cannot show: five seeds do not establish robustness; the experimental generations use different designs and scales; the ex-post benchmark is deliberately infeasible; the 200-seed holdout remains unexecuted.
Negative means lower regret for the market. Five independent simulated paths; paired SE 2.21 pp.
05 / DURABLE CAPITAL
Mises emphasized the time structure of production. This simulation tests what happens when redeploying installed capital destroys some of it—and new investment takes time.
Bars separate capital-allocation loss from other operating/coordination loss. The capital benchmark is the ex-post frictionless upper bound.
When capital becomes expensive to relocate, the best forecaster need not be the best coordinator. Investment rules, timing, and conversion costs can dominate the rankings in this particular model.
06 / WHEN AGENTS LEARN TO GAME THE SYSTEM
Experiment 4 goes beyond prediction: producers can withhold, reporters can misstate, auditors can sanction, and a learned allocation rule tries to discourage strategic deviations.
Collusion-capable producers raised markup from 0.391 to 0.422, but reduced mean welfare regret in the pilot. This goes against H2’s aggregate-surplus prediction, although true-principal utility also fell (884.89 to 868.17).
Audits and stake slashing changed mean regret from 13.554% to 13.535%—a near-null 0.020 percentage-point difference.
Mean sampled deviation gain fell from 3.4111 to 0.0268 utility units versus a simple proportional baseline.
07 / HYPOTHESES H1–H7
The hypotheses were locally documented but not independently registered. Pilot evidence may align, conflict or remain ambiguous; none is a confirmatory verdict.
08 / EVIDENCE ROADMAP
An honest research atlas should distinguish a working simulator from empirical claims about genuine agents, human participants, or payment systems.
Heuristic allocations, synthetic delegation, irreversible capital, tabular RL agents, collusion, auditing, learned Hybrid C, and discovery stress contrasts.
H1–H7, 320 factorial cells and 200 reserved holdout seeds, plus multi-vendor LLM adapters and an observed-payment calibration pipeline.
Real LLM vendor runs, consenting human principals, anonymized payment data, realistic operational costs, and independent validation.
Use the controls to count matching planned conditions. The full manifest remains downloadable and no confirmatory cell outcome is shown.
No confirmation runs claimed
09 / METHODS & SOURCES
Raw CSV files, methodology, protocol, pilot amendments and original research context are packaged with this website.