Hierarchical Skill Resolution: Enabling Skill Inheritance and Deterministic Knowledge Injection for AI Agents

Agent skills — packaged units of specialized instructions consumed by large language model (LLM) agents — are rapidly becoming the standard primitive for distributing organizational knowledge to AI systems. However, the prevailing activation mechanism is probabilistic: the model itself decides, via semantic inference over skill descriptions, whether to load a skill. This paper identifies two structural gaps that this design leaves open for enterprise adoption. First, the consumption gap: there is no guarantee that a skill is loaded when a workflow phase requires it, and no observability when it is not. Second, the inheritance gap: no current skill system supports hierarchical specialization — a global headquarters publishing base skills that country subsidiaries and business units extend with local regulatory and operational overrides. We propose the Skill Resolver, a pre-dispatch component that deterministically injects phase-mandatory skills into agent context, and Hierarchical Skill Resolution (HSR), a resolution pattern built on top of it that enables scope-based inheritance, override precedence, non-overridable invariants, and per-artifact lineage. We show how HSR composes naturally with emerging distribution infrastructure — in particular Microsoft's Agent Package Manager (APM) — forming a three-plane architecture that separates distribution, consumption, and federated governance of agentic knowledge. The pattern is targeted at multinational and multi-entity organizations, where the absence of inheritance currently forces copy-and-modify practices with predictable drift, audit loss, and linear maintenance cost.