I Built an OpenTelemetry Instrumentor for Claude Agent SDK

Table of Contents Background The problem What I built The hooks thing Getting started What the traces look like Rough edges and what's next Background A while back I wrote about going all-in on Claude Agent SDK for Holodeck . The short version: I decoupled Holodeck from Semantic Kernel through an abstraction layer, then hooked up Claude Agent SDK as a first-class backend — bash, filesystem, MCP tools, sandboxing, all native. That post was about running agents. This one is about seeing what they're doing once they run . With Semantic Kernel, this was a solved problem — it has native OpenTelemetry integration, so you got traces and metrics for free just by wiring up a provider. When I moved to Claude Agent SDK, that didn't exist. So I had to build it myself. The problem I started running longer agent sessions — multi-turn research tasks, code generation workflows, that kind of thing. And pretty quickly I realized I had no idea what was happening inside them. Like, basic stuff: How many t