Track agent performance across whole conversations, not single turns

Agent metric monitors now run at the conversation grain, not just per span or per trace. Monitor turn count, duration, total tokens, or simply how many conversations happen per bucket. Supported on any connected agent.

Cost, length, and effort are properties of a conversation, not of a single trace. An agent run that looks healthy span by span can still be the one where the user asked the same question six times.

Create an agent metric monitor here, or via the Operations agent here.