Nina: Multi-Agent Sales & Support Architecture
A multi-agent AI sales assistant that ran in production on WhatsApp for a furniture e-commerce store from July to October 2026, until the company changed its CRM setup. It has 10 specialized agents and 18 tool sub-workflows, and integrates WhatsApp, GoHighLevel CRM and Shopify.
Engineering Highlights & Architectural Decisions
CRM as Single Source of Truth
GoHighLevel had no custom AI runtime. The conversation state is serialized as a JSON object inside CRM custom fields, read and updated every turn. n8n remains entirely stateless and easily restartable.
Deterministic Rules First, LLMs Second
A 12-second debounce collects bursts of user messages. The silent L0 router (gpt-5-nano) runs strict classification rules before routing, and redirect loops are hard-capped to 3 per message.
Prices and Stock Come From Systems, Not the Model
Pricing, stock, freight and draft orders are fetched by tools from Shopify GraphQL, the store MCP endpoint and carrier APIs, and the prompts tell the agents to use them instead of guessing. This lowers the risk of wrong prices; it does not remove it.
Smooth Human Inactivity Handoff
If a customer goes quiet for 15 minutes during business hours (or 60 min at night), Nina automatically flags the contact and transfers the lead to human sales reps.