llm

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

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