Monitoring and compliance at the agent level makes sense. The gap I keep hitting is that most agent security tooling focuses on what the agent does but not what credentials it holds while doing it. An agent that passes all behavioral checks but carries a 90-day full-access API key is still one compromised session away from a bad day.