arXiv:2608.19202 formulates context acquisition as expected free energy minimization for AI agents

The paper defines an inner inference loop that maintains a posterior over latent task states and an outer policy that selects clarification, retrieval, or termination actions. It supplies an exact dynamic programming oracle for Optimal Question Asking on categorical problems ranging from 25 to 300 candidates and evaluates frontier models under explicit token budgets for both binary and multiway queries. Benchmark results show that information-gain normalized by token cost reduces clarification steps by 30-45 percent relative to unnormalized baselines while preserving task accuracy above 0.92. The same objective governs pre-generation clarification and automated prompt optimization, demonstrating consistent token savings across both regimes. Active inference supplies the normative justification that earlier clarification literature lacked; the formulation is model-agnostic and directly composes with existing agent scaffolds. It therefore supplies a measurable design criterion rather than an architectural prescription. Deployment will require tractable approximations to the expected free energy for non-tabular state spaces. Open-source implementations of the oracle and the reported model traces will appear within six months.