\section{Conclusion}
\label{sec:conclusion}

We set out to identify, inside a transparent exchange economy, the causal mechanisms through which
autonomous language-model agents change price discovery as their share of the market grows. The
instrument is an event-sourced limit-order-book simulator with exact accounting, a latent dividend
process that keeps fundamental value, public value and price apart, calibrated classical traders, and a
randomized, matched-seed design in which the LLM share is the treatment. The theory predicted that a
shared model would degrade informativeness through correlated errors, raise coordination, amplify
leverage spirals, and that diversity would help. The experiments confirmed the coordination and diversity
predictions and reversed the price predictions.

Three findings stand out. First, in this economy a small reasoning model trading from a canonical public
snapshot is, overwhelmingly, a patient liquidity provider anchored on the public present value: spreads,
impact and volatility fall and depth rises with its share, with no bubble or crash in any live run, and
with leveraged traders present the LLM agents absorb the fire sales that destabilize the classical market.
Second, the cost of that patience is slower incorporation of information that is not yet public and a
price path that becomes predictable; the variance ratio doubles at full penetration. Third, coordination
among LLM agents is high but not, in itself, harmful: the agents agree on the direction of the public
value, prompt and model diversity reduce their agreement and speed up absorption at a liquidity cost, and
a communication channel or memory raises agreement and pricing error by amounts that are precisely
estimated but small.

The honest summary of the pre-registered hypotheses is that H2 (synchrony), H4 (diversity) and H5
(communication) are supported in sign, H1 (informativeness) is supported for dynamics and rejected for
levels, H3 (instability) is rejected and its leverage corollary reversed, and H6 (safeguards) is not
testable at the shock sizes used. The stylized model that organizes the paper is a model of
market-order traders with correlated errors; the data call for a model of limit-order traders with a
common anchor. Writing that model, giving the LLM agents leverage of their own, running the Claude cells
that the design contains but a rejected key prevented, and raising the seed count to the level that
detects half-standard-deviation effects are the next steps, and the replication package makes each of
them a configuration change.

Nothing here should be read as evidence about live markets. The claim is narrower and, we think, more
useful: when the share of trading decisions taken by one language model is raised exogenously in a
market where everything else is held fixed, the first-order effect is not herding into bubbles but a
deep, calm and slow market---and the slowness, not the calm, is where the next question lies.