A trace is the tracking number for one piece of work moving through a system. It follows a single request through each stop it makes, so instead of guessing from the final answer you can see what actually happened.
Think about tracking a package. You don’t only want to know whether it arrived. You want to know when the label was created, when it reached the warehouse, when it sat at a hub, when it got loaded, and where it got delayed. A trace does that for software and AI work.
How it shows up
A trace is related to a log but not the same. A log is the warehouse’s full event book, recording many things across the whole operation; a trace is the tracking page for one request from start to finish. In agentic work, say you ask Claude Code to pull a report, clean the data, and write a summary. The agent may read files, make a tool call, call an API, hit an error, and try again. The trace is the record of that one path. It’s what lets you answer “Claude did something weird”: did the model misunderstand, did the tool return bad data, or did the job spend its time waiting on latency?
Why you care
You don’t need to read traces all day. You care when the work is slow, wrong, expensive, or risky. For clients working with AI, this is a trust issue. You don’t have to understand every technical detail, but you do need the work to be inspectable. A trace turns “something went wrong” into “here is the step where it went wrong.”