Most AI agent demos won't survive enterprise security review

OpenClaw is a genuinely impressive demo of how capable AI agents have become. It made me think about something I don’t see discussed as much: Most agent frameworks optimize for capability and iteration speed. That makes sense for personal assistants. But enterprise environments operate under very different constraints: - No inbound tunnels – Strict egress control – Identity enforcement – Tenant isolation – Audit logging – Deployment portability (local/cloud/air-gapped) When agents move from “assistant” to “worker,” the runtime layer becomes critical. Curious how others here are thinking about: What does a production-grade AI agent runtime need to look like for regulated environments? Would love to hear real-world deployment experiences.