1\section{Introduction}2\label{sec:introduction}34Trading desks, retail brokers and market-making firms have begun to connect large language models (LLMs)5to order entry. The models are few, the prompts are similar, and the information they see is the same6public tape. If a meaningful share of order flow comes to be generated by a handful of foundation models7answering near-identical questions about near-identical screens, the classical concern about correlated8trading---that it impairs the aggregation of dispersed information and amplifies shocks---acquires a9new and concrete source. Regulators have said as much \citep{fsb2024, danielsson2025}; the empirical10record on what actually happens when LLM agents populate a market is thin, because the counterfactual11market without them cannot be observed.1213This paper builds the counterfactual. We construct a transparent, replayable exchange economy in which14the share of traders controlled by a language model is an experimental treatment, and we ask six15questions: whether a rising share improves or impairs informational efficiency; whether a shared model16generates correlated orders and liquidity spirals even when prompts and private signals differ; how LLM17agents interact with fundamentalists, trend followers, noise traders, arbitrageurs, market makers and18leveraged traders; whether memory, communication and adaptive feedback improve decisions while19synchronizing them; what the causal role of agent identity, information, latency and budget is; and20which safeguards mitigate instability at what cost. The answers are causal statements about the21simulator, identified by randomization, and we are explicit that they are not statements about22institutions.2324Three design choices separate this from existing LLM market simulations25\citep{lopezlira2025trade, yang2025twinmarket, saxena2026, henning2025}. First, the economy keeps four26prices apart: the latent fundamental value defined by an exogenous dividend process whose next27innovation is drawn but not yet announced; the public value implied by announced dividends and public28news; the transaction price; and the quote midpoint. Price discovery is measured as the convergence of29the mid to the latent value, not to a known target, and information has value because the latent30innovation is revealed only through noisy news, private signals and the next dividend. Second, the31LLM treatment is randomized: a seed fixes the dividend path, the news, the endowments, the latencies and32the policy assignment of every trader slot, and the treatment replaces a stratified subset of slots by33LLM agents that inherit the information set of the slot they replace. Matched pairs differ only in who34decides. Third, the mechanisms are derived before they are tested. A stylized model of traders whose35signal errors share a model-specific component delivers closed-form comparative statics for price36informativeness, order-flow correlation, synchrony, the effective number of models, feedback37amplification and the leverage spiral; the experimental design maps each primitive to a simulator38parameter and pre-registers the sign of every prediction, the primary outcomes, the thresholds that39define bubbles and crashes and the estimators.4041The experiments use OpenAI's \code{gpt-5.4-nano} as the shared model (with \code{gpt-5.4-mini} in a42two-model and an equal-budget arm), 100-trader markets of 600 ticks, and seven matched seeds per live43cell: six shares from 0 to 100\,\% of slots, three diversity and cognition arms, a leverage interaction44with a matched classical comparator, two safeguards, two fairness arms, and a frozen-decision-library45replay at 1,000 traders. A classical design of sixteen cells and thirty seeds validates the stylized facts46and the mechanisms without any language model. The analysis is paired at the run level with Holm47adjustment within hypothesis families; with seven seeds, effects below roughly 1.2 standard deviations are48not detectable and are reported as such.4950The results separate cleanly into what the theory got right and what it got wrong. It got coordination51right: LLM agents deciding at the same tick are on the same side nine times out of ten, a communication52channel makes them more alike, and prompt or model diversity makes them less so. It got prices wrong:53raising the LLM share from 0 to 95\,\% of slots tightens the quoted spread from 18.6 to 6.1 basis points,54raises depth from 141 to 956 lots, cuts price impact by 60\,\% and leaves the deviation of price from55latent value unchanged at about half a percent; no bubble, crash or halt occurs in any of 112 live runs;56and in the presence of leveraged trend followers the LLM agents halve the pricing error and remove the57bubble episodes that leverage creates in the classical economy. The cost appears in dynamics rather than58levels: latent information is absorbed more slowly as the share rises, the five-tick variance ratio59climbs from 1.1 to 2.1 and the first-order autocorrelation of returns from 0.02 to 0.32, so that the60price reaches its value in a smooth, predictable drift that trend followers exploit. The mechanism that61unifies these facts is not correlated error but patient liquidity provision: a small reasoning model given62the public present value and asked every twenty ticks what to do mostly rests a limit order near that63value and waits. A market populated by such agents is deep, calm, and slow to learn what it does not64already know.6566Section~\ref{sec:literature} positions the paper. Section~\ref{sec:model} describes the economy,67Section~\ref{sec:theory} the stylized model and its propositions, Section~\ref{sec:identification} the68experimental design and pre-analysis plan, Section~\ref{sec:implementation} the software and its69validation. Section~\ref{sec:results} reports the classical and live experiments,70Section~\ref{sec:mechanisms} reads them against the theory, Section~\ref{sec:robustness} probes71thresholds, caching, dose, mechanism and scale, Section~\ref{sec:limitations} states what the design72cannot say, and Section~\ref{sec:conclusion} concludes. Appendices give the proofs, the prompts, the73algorithms, every hyperparameter and the replication package.74