Capability Lattice and Governance Architecture for Agentic Systems. Substrate-Governed Agent Harness Design: Threat Propagation, Ring Bounding, and the Three-Layer Stack

Agent harness architectures are commonly described in terms of surfaces—the rings of capability that compose from model to extensions. This paper moves from surface description to governance specification: what each ring is permitted to represent, what it is permitted to do, and what it is permitted to escalate to the next ring. The resulting capability lattice enables substrate-level threat bounding rather than post-hoc threat detection. The paper integrates the lattice with a three-layer conceptual model (Substrate → Harness → Runtime), a governance architecture for substrate-governed agentic systems, and a four-category threat map organized by attack mechanism rather than attack surface. The synthesis is that every major threat in agentic systems emerges when the substrate, harness, and runtime layers blur into each other.