Instrumenting Claude Code locally gets you the one thing the transcript hides, which is what the agent did between turns and how long each step took. Tool call latency and token spend per step are usually where the surprises show up. Are you capturing tool inputs and outputs as spans too, or only the timings?
Yasvanth Udayakumar
Staff Software Engineer building AI-agent infrastructure, AI observability, and cloud-native reliability systems.
Thanks, Kartik. Tempo spans capture the tool name and how long it took. The actual tool inputs and outputs are stored as Loki logs. Enabling those details can expose sensitive commands, code, and results, so I only use this setup for local development.