What a trace is
Every time the agent replies, it leaves a trace: a step-by-step, timed record of what it did to get to that reply. It’s the difference between reading what the agent said and seeing what the agent did.
What gets stored
- The turnThe agent, the conversation, the channel, what came in and what went out, and whether it ended well, with a fallback or with an error.
- The stepsEach step with its type — the agent, the model, an action —, when it started and how long it took.
- The details that matterIn the knowledge step, what was searched and which sources came up; in the memory step, which customer history was taken into account.
- The costThe model used and the input and output tokens.
What it looks like
Like a waterfall on a timeline: each step is a bar placed where it happened, indented under the step that launched it. One look tells you where the time went and what failed, and an error is marked with its icon and its label, not just a color.
What happens in tests is labeled and doesn’t mix with the real thing: a test refund is never counted as a refund made.
What your team uses it for
To understand a reply that came as a surprise, to find which step is slow, to check that the agent looked up the system before answering. And from the trace of a real case you go straight to the conversation.
