Agentic Asset Allocation and the Illusion of Evaluation: Why Multi-Agent Systems Do Not Solve the Objective Problem

This article examines the rise of multi-agent architectures in institutional asset allocation—exemplified by recent ‘self-driving portfolio’ systems—through the lens of evaluative structure. While these systems distribute forecasting, allocation, and critique across specialised agents, they operate under a fixed Investment Policy Statement (IPS) that governs selection. The central claim is that such architectures expand search over the portfolio space without addressing the specification or revision of the objective function that determines outcomes. The analysis formalises a distinction between optimisation and evaluation. In current systems, the evaluative structure remains external to system operation, so forecast errors and model disagreement are processed as failures relative to a fixed objective rather than as evidence of objective mis-specification. Reinterpreting multi-agent systems as ensemble optimisers, the article shows that diversity of models does not imply independence of outcomes when evaluative criteria are shared. A stylised simulation isolates the consequences of shared evaluation. Increasing similarity of evaluative structure produces greater portfolio overlap, higher aggregate concentration, and stronger drawdown correlation under stress. These effects persist under partial overlap, scale with the degree of similarity, and are only partially mitigated by endogenous regime switching, which itself occurs in a clustered manner. The article makes five contributions. It defines the objective problem in financial AI as the absence of mechanisms for specifying and revising evaluative criteria; establishes a formal taxonomy distinguishing parametric adaptation from evaluative revision proper; introduces evaluative crowding as a source of convergence and fragility; provides model-bounded evidence that evaluative similarity maps into concentration and stress amplification; and proposes Objective-Layer AI as a minimal architectural extension in which objectives become admissible objects of constrained transformation. The next frontier in institutional asset management lies not in scaling optimisation, but in designing and governing the objective layer itself.