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

When AI agents become economic agents, the economy acquires a new source of aggregate fluctuations that has nothing
to do with technology shocks: the mistakes of a few shared models, acting through millions of delegated decisions.
This paper has tried to say precisely what that means and, just as precisely, what it does not.

It means that a market for delegation sets the error-covariance structure of the economy. The delegation share and
the effective number of independent models are sufficient statistics for how much capital, wages, interest rates and
asset prices load on model noise, and the loading has a closed form. It means that whoever owns the compute that
delegated decisions run on captures a new kind of rent, and that better decisions are themselves a scale-dependent
return that concentrates wealth. And it means that markets populated by agents sharing one model are quieter and
more correlated at the same time, so that their fragility shows up conditional on a false shared signal, through
collateral, and depends on how execution turns correlated orders into a marked price.

It does not mean that delegation is inefficient. In the benchmark economy the private and social values of
delegation coincide exactly; the externality that policy discussions take for granted exists only under
identified frictions, and we have given its size and sign under each. Model diversity is a policy instrument only
where those frictions bind; a delegation tax and a diversity mandate are substitutes there, and both are
unnecessary where they do not. Nor does it mean that language-model agents behave like people: our laboratory shows
what the architecture does to a market, under constraints that no model can override.

Three extensions follow directly from the code. A Krusell--Smith version of Model B would put the delegation game
inside each period of an economy with aggregate risk and let the decision-quality channel and the model-noise channel
interact. A firm-side version would let managers delegate input choices and quantify the productivity effect of
advisory AI through reduced misallocation. And the empirical roadmap of Section~\ref{sec:calibration}---adoption,
transaction and broker data linked to the identity of the model behind each agent---would turn the effective number
of models from an illustrative parameter into a measured one. The last is the most important: the quantity that
matters for the economy is not how many foundation models exist, but how many of them make independent mistakes.