New Primitives Proposed for Governing Autonomous AI Agents at Runtime

Traditional control models, designed for human users and long-lived services, are ill-suited for governing autonomous AI agents in enterprise environments. The unique characteristics of agents—their ephemeral existence, model-selected actions, and dynamic, discovered populations—create significant governance gaps. This research posits that governing these agents is fundamentally a runtime problem, distinct from model alignment or build-time concerns. The authors derive five essential primitives necessary for effective runtime governance: discovery, identity, governance, attestation, and supply chain. Each primitive addresses a critical failure point if absent, and they are structurally interdependent. Discovery ensures agents can be identified, identity provides a verifiable persona, governance mediates actions against policy, attestation records actions for verification, and supply chain tracks the provenance of agent components. An implemented architecture demonstrates these primitives, where agent actions are mediated against policy before execution, authorized by a per-tenant action vocabulary, and recorded in a verifiable, hash-linked ledger. While effective, this approach introduces costs: enforcement points are on the critical path, identity requires sidecars, and fail-closed mediation can convert availability issues into denial-of-service. The work emphasizes that four primitives are operational in private pilots, with the fifth (supply chain) as separate tooling, highlighting the practical challenges and ongoing development in this crucial area.