1Households and firms are beginning to delegate economic decisions to autonomous agents built on a few shared2foundation models. We study what this does in general equilibrium. A human's decision error is idiosyncratic and3averages out; an AI agent's error is common to every principal that uses the same model. Delegation therefore4converts noise that cancels into noise that does not: capital accumulation, asset demand and credit load on the5shared error with weight $D$ (the delegation share) and $1/M$ (the effective number of independent models), and6wages, interest rates and asset prices inherit this ``model noise''. In a task-based stochastic growth economy with a7compute market, the decision problem reduces to a linear-quadratic game with dispersed information in which8delegation is the purchase of a correlated signal. We prove that in a price-taking economy with constant returns the9private value of delegation equals its social marginal value identically---model noise is real but not an10inefficiency---and we obtain closed-form wedges, Pigouvian delegation taxes and the rate at which model diversity11substitutes for them when the aggregate action has non-pecuniary effects, when price risk is uninsurable, or when12prices enter collateral constraints. Delegating the consumption--savings decision has almost no aggregate bite;13delegating the allocation of wealth has a large one. In a heterogeneous-agent economy the rich delegate first, better14decisions concentrate wealth even with broad ownership of AI, and the distributional effect of automation is decided15by who owns the compute. In an agentic market laboratory with explicit balance sheets, margin rules, defaults, three16market designs and live open-weight language-model agents under hard deterministic constraints, order-flow17correlation rises with $D$ and falls with $M$ in every design; a shared model makes markets quieter and more18correlated at once; and a false shared signal is amplified by collateral in a call market (by a factor of 2.6 at a1920\% margin, halved at 50\%, eliminated by model diversity) but not in a continuous auction. Language-model agents herd less than Bayesian agents on the same shared signal but load prices on it more, and their20proposed actions routinely violate the constraints that deterministic code enforces. Code, data provenance, prompts21and audit logs are provided.22