What is an agent trace?
A trace is the step-by-step record of one run: what the model saw, which tool it called, and what came back. An audit log says who approved or sent. A trace says how the draft was built. You need the trace when the words are strange.
Also called: run trace
In plain English
Open the run and see that it pulled the wrong guest's bio, then wrote a lovely email. The loveliness is not the bug. The tool call is. If the product does not show tool calls, you are guessing. Read the trace like a receipt. Which document came back, which id was blank, which tool was skipped. Fix that line. A trace can contain the customer's own words and any secret you foolishly passed as an argument, so do not forward the raw thing to a group chat or a second model. If the product hides tool results, you will guess, and the guesses will be about the prompt when the bug was the data. Compare two traces of the same job, one good and one bad. The first step that differs is the one to repair, not the final paragraph. Save both until the fix sticks. Archive the healthy run beside the broken one through a Monday.
How an owner would use it
When a draft is wrong, open the trace before you rewrite the prompt at random. Look for the tool result. If the tool returned the wrong row, fix the tool or the id. If the tool was right and the draft ignored it, fix the prompt.
Common mistakes
Editing the prompt every time, without reading the trace, thrashes a workflow that was fed bad data. Traces can contain customer content. Do not forward a raw trace into another AI tool. Do not keep secrets in tool arguments if the trace is visible to the whole team.
A worked example
A small podcast studio
The trace shows the wrong episode notes
This is an illustration, not a customer result. A guest gets a prep email that mentions a different show. The trace shows a search that returned last month's notes because the episode id was blank. They fix the id in the trigger. They do not add more adjectives to the prompt.
Related terms
Prompts and guides
Questions owners ask
What is a trace?
A trace is the ordered detail of one agent or workflow run, including tool calls and their results.
How do I use a trace to fix a bad draft?
See which input and which tool result the draft used. Correct that source before you rewrite the instructions.
How is a trace different from an audit log?
The audit log is for accountability: who sent or approved. The trace is for diagnosis: which step went sideways.
Product guides are on AI apps for business.