\section{Limitations}
\label{sec:limitations}

We list the limitations that a referee should weigh, in decreasing order of importance.

\paragraph{The efficiency benchmark is a second-order result in a price-taking economy.} Proposition~\ref{prop:benchmark}
holds in the linear-quadratic approximation around a symmetric steady state with complete markets. Higher-order
terms, market power of model providers, and non-pecuniary spillovers (knowledge, networks, data externalities) all
break it, in the direction given by Proposition~\ref{prop:wedges}(a). The benchmark is a reference point, not a claim
that delegation is efficient in practice.

\paragraph{Which decision is delegated is assumed, not derived.} The portfolio mapping is our baseline because the
savings mapping has no bite. A fuller treatment would let households delegate a menu of decisions (allocation,
labour supply, consumption timing, firm input choices) with heterogeneous stakes; the firm-level misallocation channel
of \citet{david2016information} is the natural next extension and would add a productivity effect of advisory AI that
we do not quantify.

\paragraph{Information parameters are illustrative.} The precision of human and model forecasts of returns, the
effective number of independent models, the supervision time and the delegation disutility have no accepted
empirical counterparts. We sweep them widely and report which conclusions are robust, but the \emph{levels} of
welfare effects in Section~\ref{sec:results_a} should not be quoted as estimates. The same applies to the Model B
return-noise parameters.

\paragraph{Model B abstracts from aggregate risk.} The stationary equilibrium cannot exhibit model noise; it isolates
the cross-sectional channels (who delegates, who owns). A Krusell--Smith extension with the delegation game inside each
period is feasible with the present code and left for future work. Like all Aiyagari-type models without additional
return heterogeneity, Model B understates the top-1\% wealth share.

\paragraph{Model C is a laboratory, not an equilibrium.} Outcomes are bounded-rationality simulated outcomes under a
temporary trading equilibrium per period. The market maker's ladder is a stand-in for liquidity provision; depth and
price impact are parameters, not equilibrium objects. The number of agents is small (20 in the control, 8--10 with
live models), so order-flow correlations and price-pressure $R^2$ are noisy; we report bootstrap intervals and treat
the LLM results as qualitative.

\paragraph{LLM agents are not human subjects.} The models used are small open-weight instruction models at 4-bit
precision; their behaviour depends on prompts, temperature and version, which we vary and log. Nothing in the LLM
results is evidence about human preferences or about how deployed commercial agents behave; the results show what
the \emph{architecture} (shared versus diverse models, with hard constraints) does to a market populated by such
agents.

\paragraph{No external causal identification.} All effects are identified within the structural model by
construction. The external-validation roadmap in Section~\ref{sec:calibration} lists the data (adoption, transaction,
13F, broker records) that would be needed to test H1--H6 outside the model.

\paragraph{Legal and institutional scope.} The baseline treats AI agents as policies with human or corporate
principals. The independent-AI-principal counterfactual is implemented in the code (an investor with no residual
claimant) but is not reported, because we found no institutional benchmark against which to calibrate its balance
sheet or its liability.