Building Loop Agent: One Worker, Every Channel

The pattern I kept running into across client work — BYD, Aga Khan, and a handful of smaller builds — was the same failure mode: a voice bot that answers calls well, a chatbot that handles WhatsApp fine, and no shared memory between them. A customer calls, gets qualified, then messages on WhatsApp an hour later and starts from zero. The business paid for two AI tools and got one broken conversation.

Loop Agent is my answer to that: one agent, one context object, however many channels the customer actually uses.

Instead of building a voice bot and a chat bot as separate systems that occasionally sync, Loop Agent treats every channel as a different transport for the same underlying conversation thread. Voice, WhatsApp, and follow-up messaging all read from and write to one context record tied to the customer — intent, history, CRM state, and calendar availability all live in one place, not three.

Concretely: if a customer starts on a call and the conversation ends in 'let me think about it,' the follow-up message that goes out later isn't a generic nudge — it's grounded in exactly what was discussed on the call. If they then reply on WhatsApp, that reply lands with full context, not as a fresh inbound message from a stranger.

Right now Loop Agent is an internal product, in development — not a deployed client project. The architecture draws directly from patterns proven in live client work: Vapi for voice orchestration, Telnyx for number provisioning, Supabase with row-level security for the shared context store, and Cal.com for scheduling. The multi-tenant piece — modeling each client as a squad of specialized agents (reception, sales, follow-up) with mid-call handoffs — is the part carried over most directly from the BYD build.

What's still ahead: hardening the handoff logic for edge cases, and getting the first non-internal deployment live so the 'one context, every channel' claim gets tested against a real, messy customer base instead of internal test scenarios.

It would be easier to sell Loop Agent as a chatbot with a voice add-on. But that's the exact framing that produces the disconnected-tools problem in the first place. The bet here is that the hard part isn't answering a call or replying to a WhatsApp message — plenty of tools do both. The hard part is making sure the business doesn't lose the thread the moment a customer changes channels, which is also the moment most AI tooling quietly breaks.