AI Control Tower assumes you know what agents exist. Most instances I've looked at don't have that list anywhere
I built a read-only scanner to test that assumption, agents, tools, and credentials already running on an instance, correlated by confidence rather than reported as a flat yes or no. One scan found seven agents that were never registered as agents anywhere, scripts and flows quietly calling out to a model, plus a generative AI credential with no resolvable owner.
Detection is rule based and deterministic, so the same instance produces the same findings every run, which matters if the output is ever going to sit in front of an auditor. Verified against three live instances.
Curious how others here are handling the inventory question before AI Control Tower policy even applies, is anyone auditing what's actually running versus what's registered?
Writeup: https://www.linkedin.com/pulse/you-cant-govern-ai-your-instance-dont-know-its-alex-mcdonald-mllve
Repo: github.com/BrianMcD47/AgentCensus
