The single chatbot from v0.1 didn't scale to the kinds of questions our early users were asking. v0.2 introduced the multi-agent architecture that still underpins Curated today.
What shipped
- Google ADK (Agent Development Kit) backend — a new FastAPI service (then internally called Bumblebee AI) hosting purpose-built agents
- GIS Agent — handles geocoding, layer queries, spatial analysis, routing
- Salesforce Agent — pulls account and opportunity data and grounds answers in customer records
- Real-time streaming — answers arrive token-by-token instead of after a long wait
- WebSocket session management with heartbeat + exponential backoff reconnection — production stability from the start
Why it mattered
A single LLM trying to do GIS, CRM, and conversational reasoning all at once gets each of them wrong. Specialists do better. v0.2 set up the orchestration so a question like "Which Salesforce accounts are within 10 miles of our new Houston location?"* could route the GIS half to one agent and the CRM half to another, then compose the answer. The pattern is the same today — we just have eight specialists instead of two.