Post-IAM Governance: Why Agent Registries Are Necessary but Insufficient for Agentic AI Accountability

The emergence of enterprise agent registries — systems that catalog, track, and govern AI agents through metadata, invocation logs, and approval workflows — represents a meaningful advance in organizational AI management. This paper argues, however, that agent registries reproduce the governance assumptions of Identity and Access Management (IAM) — a paradigm designed for human actors and service accounts operating in deterministic, bounded systems — and apply them to agentic AI systems that systematically violate every assumption IAM was built on. IAM-era governance assumes stable identity, bounded permission sets, predictable action scopes, and logs that accurately reflect what happened. Agentic systems violate all four assumptions: identity mutates across sessions, privilege envelopes drift through delegation chains, sub-agents act outside declared orchestrator scope, and agent-generated logs reflect reported state rather than ground-truth reasoning. An agent registry that tracks metadata, publication, invocation details, and approval workflows — however well-implemented — cannot see planner-layer reasoning, intent deformation, privilege envelope drift, sub-agent lineage, behavioral invariants, or identity mutation. These are not implementation gaps. They are structural limits of the IAM paradigm applied to the wrong class of actor. This paper defines post-IAM governance — the architectural requirements for governing agentic systems that operate beyond the IAM visibility horizon — and describes the substrate-layer infrastructure that closes the gap registries cannot close. The APR-Series governance stack (APR-Shaper, APR-Guard, APR-Lite), developed by Soft Armor Labs, is presented as a reference architecture for post-IAM governance.