One prompt, three backends: a Postgres semantic layer, a Pinecone document index, and a read-only SQL escape hatch. Tool calls stream inline so you can see exactly what the agent did.