The short answer
An AI-agent pilot should prove more than a successful conversation. Test data scope, identity, sandbox enforcement, privileged requests, resource ceilings, update recovery, evidence retention and a real shutdown path. Compare Clawsolator with native OpenClaw and NVIDIA tools using the same scenarios, installed versions and operator goals.
Identity and authority: six questions
Who owns the gateway credentials? How is an individual agent identified? Can tokens be rotated or revoked? Can a requester confer greater privileges on itself? Which actions demand a human? Can approvals be reviewed after a restart?
Sandbox and network: six questions
Can an agent read another team’s files? What happens on denied egress? What filesystem enforcement exists? Can a process access the host socket? Are the default policies documented? Is a real denial test recorded?
Operations and costs: six questions
Who sets CPU and RAM caps? Can one agent exhaust the host? Where are change events written? Can you enroll and retire an agent predictably? What monitoring survives reboot? Can the CFO attribute costs to a workload?
Recovery and governance: six questions
What is backed up? Where are secrets stored? Can state be restored on another host? How are policies verified after updates? Who resolves a failed approval? What are the kill-switch and incident escalation paths?
Vendor/product diligence: six questions
Which features are shipping today? What is documented as planned? What is the exact test environment? Who controls licensing and distribution rights? How portable are data and configurations? Can the provider show reproducible results rather than screenshots?
Make the responses comparable
Score every item as demonstrated, documented but not reproduced, planned, or absent. Attach a command/output capture, version and owner to each demonstrated item. That is how a technically skeptical executive can feel comfortable recommending a pilot.
Questions people actually ask
Is one green compliance score enough?
No. Require the specific control and evidence relevant to the deployment threat model.
Can I use this checklist with NVIDIA?
Yes. It is intentionally vendor-neutral so a buyer can compare OpenShell, NemoClaw, OpenClaw and Clawsolator with the same tests.
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.