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 you can see what actually happened instead of guessing from the final answer. A log records everything across the operation; a trace follows one request from start to finish.
Imagine 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, and where it got delayed. A trace does that for one piece of AI work.
Why you care
Ask an agent to pull a report, clean the data, and write a summary, and it may read files, make a tool call, hit an error, and try again. The trace records that one path, so “Claude did something weird” becomes answerable: the model misunderstood, the tool returned bad data, or the job sat waiting on latency. You won’t read traces all day; you care when work is slow, wrong, or expensive. For clients this is a trust issue: the work needs to be inspectable.