Closing the loop: using coding agent traces in Pydantic Logfire

Coding agents do real work every day, and almost none of it gets recorded. The commit and the PR survive, but how the agent got there doesn't. When a result is bad, you have few good options. You can ask an agent to fix it without showing it where things went wrong. You can rewrite your instructions without knowing what was off. Or you can switch tools without knowing what you're giving up.

Every question you'd want to ask is answered in the traces. Which harness is best at code review? Would a cheaper model give the same results? Did the agent get confused by its tools and burn two million tokens? Is it re-exploring the same tasks every session?

In this workshop, Duncan McKinnon sets up tracing for a coding harness into Pydantic Logfire. He walks through what a trace holds, from the full session down to each turn, model call, tool run and subagent. Then he puts that OpenTelemetry data to work by querying it with the Logfire MCP server, and builds the metrics you should use to evaluate your harness.