Every routing decision is a model call.
Three specialist RAG agents split the Larkspur Cloud handbook: billing, operations and data. Send a request through a router, a supervisor, a hierarchy or a swarm, add follow-up turns on the same thread, and count the model calls. Magenta rows are calls spent on deciding who works.
- Model calls, whole thread
- Spent on routing
- Specialists reached, first turn
- Calls per follow-up
Call trace
| Call | Node | Inside | What the call does |
|---|
All four at this setting
| Architecture | Model calls | Routing calls | Reached |
|---|
Each node is one model call: a routing decision or a specialist’s answer. Retrieval tool calls are left out, because every architecture makes the same ones. The traces match LangGraph graphs with the same nodes, run with stub models.