observability

One message, dozens of calls: how we measured and cut agent-chat spend

A chat message is not one request — it is a loop of model calls. How we attributed LiteLLM spend, saw the whole proxy in Grafana, and cut tokens without breaking the agent.