Enterprise deployments of autonomous AI agents inherit a control model originally built for human users and long‑lived services, yet it mismatches in three ways: the agents’ principals are fleeting, appearing and disappearing faster than provisioning can keep up; their actions are chosen by a model rather than hard‑coded, so the set of possible behaviors cannot be enumerated in advance; and the agent population is discovered via API calls instead of being provisioned ahead of time. We argue that governing such agents is a runtime problem, not a model‑alignment or build‑time issue. From this we derive five essential primitives: discovery, identity, governance, attestation, and supply‑chain. The absence of any primitive leads to system failure, and the others cannot structurally replace it. In our implementation, each agent action is mediated against policy before it takes effect, authorized against a per‑tenant action vocabulary, and recorded in a hash‑linked signed ledger that third parties can verify without involving the vendor. Architectural costs appear as follows: the mediation point sits on the request’s critical path, identity requires a sidecar per workload, and fail‑closed mediation turns availability incidents into denials. Four primitives are already built and running in private pilots; the fifth is provided as separate tooling and has not yet been integrated into the request path. We stress that this five‑part decomposition is not a taxonomy but a description of the actual codebase.
Blogger's Review: The paper neatly isolates the runtime governance needs of autonomous AI agents into five concrete modules, highlighting the shortcomings of traditional control models while offering a practical implementation blueprint. This approach gives enterprises a clear path to enforce policy, maintain auditability, and ensure verifiable execution of AI‑driven actions.