Running a WhatsApp agent in production is an operations problem wearing an AI costume. The first message is easy. The ninetieth day is the test.
Upkeep, ranked by how often it bites
- Reconnection after a dropped link — and alerting the user, because a silent disconnect looks exactly like a quiet week.
- Media handling: voice notes, images and PDFs arriving faster than you transcribe them.
- Timezone and DST correctness for every scheduled thing you own.
- Cost drift as chat volume grows, especially for anything that classifies every inbound message.
- Model and dependency upgrades that quietly change behaviour your rules relied on.
The guardrails you cannot add later
Bulk-send caps, bulk-delete caps, and outright refusal of mass group-adds have to be in the send path from the first commit. Bolted on afterwards, they are always circumventable — usually by the agent itself, splitting the work across turns to be helpful.
Cheap gates before expensive models
Do not send every message to a large model. Match exact triggers with plain code, use a small classifier as the gate for anything requiring judgment, and reserve the expensive path for work that has already earned it. This is the difference between a hobby bill and a business one.
Observability without surveillance
If you promise no activity logs, you cannot debug by reading conversations. That constraint is workable: keep structured, non-content telemetry on the control plane only, and make debugging inside the user's environment something the user's own agent performs on request.