What’s your approach to end-to-end AI cost attribution (model + infra + data) for agents in production?

Cost attribution gets even worse when you factor in the bad decisions. Retries, abandoned tool calls, and cascading failures that still hit your bill but never show up in traces.