The short answer
Keep requesters and approvers on different authority paths. An agent should be able to describe a needed host operation without automatically receiving sudo, an unrestricted shell, or the container engine. XMAN’s daemon has separate requester and operator interfaces; sensitive actions are submitted for review, with durable identifiers and bounded execution policies.
The sentence that should stop a demo
“The agent needed the permission, so I gave it admin access” is not a permission model. Separate the request, the decision, and the execution. Ask what exact script is being approved, under which identity, with what effect and rollback plan.
OpenClaw provides tool policy and approvals; OpenShell also provides policy controls. The contrast is the broader workflow around host-level exception requests, preserved records, and human review across an evolving installation.
The XMAN approval path
The XMAN v0.4.76a1 architecture documents per-agent scoped requester tokens, private Unix-socket APIs, Linux peer-identity checks, typed operations, operator review, and event receipts. An operator can inspect a proposed special action before it is applied. That is distinct from exposing a generic sudo bridge to an LLM.
Read approval workflows and daemon boundaries. The redacted recording is an owner-supplied demonstration; it is not an external penetration test.
A reproducible acceptance test
Submit a safe request that should be rejected, another that needs approval, and a bounded resource operation. Verify the event IDs, actor scope, decision, exact executed command, and readback after execution. Reboot during a pending request and inspect recovery behavior on the actual release, rather than inferring it from a presentation.
What an executive should hear
The product is not claiming that every prompt is harmless. It is arguing that the prompt cannot confer new host authority on itself. That is a procurement question, an operations question, and a security question in one.
Questions people actually ask
Do native OpenClaw approvals make XMAN unnecessary?
Possibly for a simple deployment. Compare your required human approval, persisted evidence, resource management, and host-change flows rather than treating them as identical or mutually exclusive.
Is a Unix socket enough to secure a daemon?
No. Socket permissions, peer identity, token handling, operation scope, input validation and the host trust model all matter.
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.