Dochkina (2026) reports a striking empirical finding: among multi-agent LLM systems, a hybrid Sequential protocol, in which agent ordering is fixed but role selection is fully endogenous, outperforms both centralized coordination (+14%, p < 0.001) and fully autonomous protocols (+44%, Cohen's d = 1.86, p < 0.0001). The paper names this the endogeneity paradox but does not offer a theoretical mechanism explaining why completed outputs from predecessors are informationally superior to intentions, historical patterns, or a central plan. This response supplies that mechanism. Drawing on the Generalized Intentionality Mismatch Theorem (Lerer, 2025a), the theory of common knowledge as an epistemic device (Pinker, 2025; Aumann, 1976), and the Extended Phenotype Theory applied to institutional design (Lerer, 2025b), this paper argues that the Sequential protocol succeeds because it is the minimal institutional structure capable of generating common knowledge among agents with heterogeneous intentionality, without requiring those agents to model each other's mental states. The result generalizes beyond LLM systems to any multi-agent coordination problem, including legal institutions, collegial tribunals, bicameral legislatures, and procedural sequences in both common law and civil law traditions, where the same epistemic condition governs equilibrium stability.
Ignacio Adrián LERER (Fri,) studied this question.