LaunchDarkly now stitches related LLM and agent spans into conversations, so a run that spans many traces reads as one transcript instead of a pile of spans. Open Traces and select the Conversations tab to see duration, models, providers, token usage, errors, and estimated cost for the whole run, then read the transcript turn by turn, including tool calls and their payloads, with evaluation scores shown on the turn they describe.

To get started, read LLM observability and Instrumenting LLM applications.

Our team is happy to help you instrument your application. To get help or share feedback, contact your LaunchDarkly account team or email product@launchdarkly.com.