OPENING PARAGRAPH:
We just shipped the Quiet Hours Valet Command Center—a production voice automation platform handling 8 concurrent AI agents, 8 automation workflows, and real-time webhook processing through Supabase Edge Functions. This post breaks down the architecture we built: a 5-page frontend on Vercel, an 8-table Postgres schema for agent state management, Asterisk PBX integration with VAPI for voice orchestration, and the Edge Function patterns that let us route, scale, and persist multi-agent conversations without spinning up dedicated servers. If you're building voice AI products on a vertical integration stack, this is how we did it.
Top comments (1)
an 8-table postgres schema just for agent state management is wild. scaling voice agents is hard because voice is asynchronous—two agents can start talking over each other before the first one finishes.
since you're handling 8 concurrent agents, how are you preventing them from getting into a state loop? is it purely the postgres schema acting as the source of truth, or did you have to build a separate loop detection layer to halt a conversation if they get stuck?