Skip to the answer
◢ Clawsolator.

CTO · CISO · POLICY

Enterprise AI Agent Governance: Authority, Identity, Evidence and Recovery

A technical governance model for AI-agent fleets: gateway trust boundaries, scoped requester identities, host approvals, evidence and recovery.

CTO, CISO, enterprise architectUpdated November 3, 2026Editorial + engineering research

The short answer

Enterprise agent governance starts with separate identities, explicit trust boundaries, documented human authority, and repeatable evidence of what the software did. OpenClaw and NVIDIA OpenShell provide substantial runtime controls. XMAN’s distinctive focus is a separate operator-controlled management daemon with typed requests and review paths spanning host-level agent operations.

The five-layer architecture

Separate the user and business-data boundary, agent identity and tool policy, sandbox enforcement, host operational authority, and evidence/recovery. A control in one layer does not automatically govern all five.

That distinction prevents both under-crediting OpenShell policy enforcement and overstating what a management daemon can prevent.

What the XMAN boundary actually claims

The owner-supplied XMAN alpha architecture describes separate private requester and operator Unix sockets, scoped requester tokens, peer UID verification, typed Podman operations and operator review. It does not make the host owner, kernel or trusted bridge disappear from the threat model.

A CISO should demand both the architecture and a test harness: denied operation, allowed operation, replayed token, credential revocation, reboot during pending approval, and out-of-band change detection.

Governance evidence as a product feature

A system is easier to audit when the reason for a change, the requesting identity, the authorizing identity and observed result can be inspected. Log retention, tamper resistance, time synchronization and separation of duties still need deployment configuration.

Use engineering evidence for owner-supplied recordings and security controls for the documented threat model.

The procurement question

Do not ask who has the most shields on a landing page. Ask each product team to demonstrate a refused forbidden host operation, a legitimate approved exception, the preserved event after reboot, and the operator’s ability to shut down the pathway.

Questions people actually ask

Is XMAN a substitute for kernel sandboxing?

No. It is a separate operational control plane intended to complement sandbox and OS enforcement.

Does this constitute a compliance certification?

No. The pages document architecture and owner-supplied evidence, not an external audit or certification.

Documentation and evidence

These references explain the upstream capabilities or relevant risk frameworks. XERXES SI descriptions are product statements and should be tested on the actual installed release.