How are you enforcing runtime policy for AI agents today?

Over the last year, I have been building and integrating AI agents into real systems and keep hitting the same concern: once an agent can call tools and APIs, how do we enforce what it is allowed to do, not just log what it did after the fact?

I am curious how others are handling this. Do you use policy-as-code, allowlists, approvals, or something else to govern agents at runtime? What controls or evidence would make you comfortable letting agents touch production data and systems?

I am working on Execlave, AI Agent Governance & Enforcement Platform (runtime AMP) that sits as a gate in front of agents, but I would really like to hear how you are approaching this problem today and what is missing.