Heterogeneous Compute Routing: Capacity-Aware Multi-Policy Workload Placement

Production agentic operations route workloads across heterogeneous substrates: on-premises clusters, hyperscaler-managed inference surfaces, sovereign bare-metal capacity, and forward-deployed edge nodes. The naive view of routing in such environments treats workload placement as a load-balancing problem to be solved at the orchestration plane. This paper argues that routing in heterogeneous compute is a capacity-aware multi-policy decision that reasons simultaneously about workload class, governance envelope, substrate availability, cost regime, and sovereignty constraint. Single-policy routers (latency-only, cost-only, region-only, SKU-only, account-only) fail in production agentic operation because the binding constraint at any given moment may originate in any of the five policy dimensions, and a router that optimizes one dimension while ignoring the others is operating outside the federation envelope without recognizing it. The paper develops a specification of the capacity-aware multi-policy router and shows how it composes with the K3s + service mesh substrate that the Mission-Ready Decalogy's operational plan documents in production. It identifies five structural failure modes of single-policy routing, specifies the multi-policy decision logic, and binds the routing layer to the Skipjack Protocol's enforcement mechanisms. Routing policy is governance-defined and human-owned; routing decisions within policy are agentic. This is Paper 2 of the *Managing Agentics Ops* series. Paper 1 (Capacity Sourcing) establishes the six-layer substrate; Paper 3 (Capex/Opex Inflection) examines the cost-regime conditions that determine when routing across substrates beats consolidation on any one.