\section{Introduction}
\label{sec:introduction}

Trading desks, retail brokers and market-making firms have begun to connect large language models (LLMs)
to order entry. The models are few, the prompts are similar, and the information they see is the same
public tape. If a meaningful share of order flow comes to be generated by a handful of foundation models
answering near-identical questions about near-identical screens, the classical concern about correlated
trading---that it impairs the aggregation of dispersed information and amplifies shocks---acquires a
new and concrete source. Regulators have said as much \citep{fsb2024, danielsson2025}; the empirical
record on what actually happens when LLM agents populate a market is thin, because the counterfactual
market without them cannot be observed.

This paper builds the counterfactual. We construct a transparent, replayable exchange economy in which
the share of traders controlled by a language model is an experimental treatment, and we ask six
questions: whether a rising share improves or impairs informational efficiency; whether a shared model
generates correlated orders and liquidity spirals even when prompts and private signals differ; how LLM
agents interact with fundamentalists, trend followers, noise traders, arbitrageurs, market makers and
leveraged traders; whether memory, communication and adaptive feedback improve decisions while
synchronizing them; what the causal role of agent identity, information, latency and budget is; and
which safeguards mitigate instability at what cost. The answers are causal statements about the
simulator, identified by randomization, and we are explicit that they are not statements about
institutions.

Three design choices separate this from existing LLM market simulations
\citep{lopezlira2025trade, yang2025twinmarket, saxena2026, henning2025}. First, the economy keeps four
prices apart: the latent fundamental value defined by an exogenous dividend process whose next
innovation is drawn but not yet announced; the public value implied by announced dividends and public
news; the transaction price; and the quote midpoint. Price discovery is measured as the convergence of
the mid to the latent value, not to a known target, and information has value because the latent
innovation is revealed only through noisy news, private signals and the next dividend. Second, the
LLM treatment is randomized: a seed fixes the dividend path, the news, the endowments, the latencies and
the policy assignment of every trader slot, and the treatment replaces a stratified subset of slots by
LLM agents that inherit the information set of the slot they replace. Matched pairs differ only in who
decides. Third, the mechanisms are derived before they are tested. A stylized model of traders whose
signal errors share a model-specific component delivers closed-form comparative statics for price
informativeness, order-flow correlation, synchrony, the effective number of models, feedback
amplification and the leverage spiral; the experimental design maps each primitive to a simulator
parameter and pre-registers the sign of every prediction, the primary outcomes, the thresholds that
define bubbles and crashes and the estimators.

The experiments use OpenAI's \code{gpt-5.4-nano} as the shared model (with \code{gpt-5.4-mini} in a
two-model and an equal-budget arm), 100-trader markets of 600 ticks, and seven matched seeds per live
cell: six shares from 0 to 100\,\% of slots, three diversity and cognition arms, a leverage interaction
with a matched classical comparator, two safeguards, two fairness arms, and a frozen-decision-library
replay at 1,000 traders. A classical design of sixteen cells and thirty seeds validates the stylized facts
and the mechanisms without any language model. The analysis is paired at the run level with Holm
adjustment within hypothesis families; with seven seeds, effects below roughly 1.2 standard deviations are
not detectable and are reported as such.

The results separate cleanly into what the theory got right and what it got wrong. It got coordination
right: LLM agents deciding at the same tick are on the same side nine times out of ten, a communication
channel makes them more alike, and prompt or model diversity makes them less so. It got prices wrong:
raising the LLM share from 0 to 95\,\% of slots tightens the quoted spread from 18.6 to 6.1 basis points,
raises depth from 141 to 956 lots, cuts price impact by 60\,\% and leaves the deviation of price from
latent value unchanged at about half a percent; no bubble, crash or halt occurs in any of 112 live runs;
and in the presence of leveraged trend followers the LLM agents halve the pricing error and remove the
bubble episodes that leverage creates in the classical economy. The cost appears in dynamics rather than
levels: latent information is absorbed more slowly as the share rises, the five-tick variance ratio
climbs from 1.1 to 2.1 and the first-order autocorrelation of returns from 0.02 to 0.32, so that the
price reaches its value in a smooth, predictable drift that trend followers exploit. The mechanism that
unifies these facts is not correlated error but patient liquidity provision: a small reasoning model given
the public present value and asked every twenty ticks what to do mostly rests a limit order near that
value and waits. A market populated by such agents is deep, calm, and slow to learn what it does not
already know.

Section~\ref{sec:literature} positions the paper. Section~\ref{sec:model} describes the economy,
Section~\ref{sec:theory} the stylized model and its propositions, Section~\ref{sec:identification} the
experimental design and pre-analysis plan, Section~\ref{sec:implementation} the software and its
validation. Section~\ref{sec:results} reports the classical and live experiments,
Section~\ref{sec:mechanisms} reads them against the theory, Section~\ref{sec:robustness} probes
thresholds, caching, dose, mechanism and scale, Section~\ref{sec:limitations} states what the design
cannot say, and Section~\ref{sec:conclusion} concludes. Appendices give the proofs, the prompts, the
algorithms, every hyperparameter and the replication package.