Hey - Really clean architecture on the outbound-only relay — solving the NAT problem that way is elegant.

Curious how you’re thinking about observability once agents are actually running. You can see which agent handled a message and where, but do you get any visibility into what happened inside the session — like reasoning steps, tool calls, token usage per convo?

The privacy routing layer is super compelling, but I’d imagine teams putting this into production would want that inner visibility too — especially for cloud agents where you’re effectively trusting a third party with execution.

How are you thinking about debugging when a cloud agent gives an unexpected response?

lol did you just comment on your own AI thread with more AI slop?

[dead]