Tracing

Every run of a deployed agent is traced automatically. Traces are how you see what an agent actually did — step by step, with the inputs, outputs, and cost of each run — and they're collected in your own account.

What a trace records

A run is captured as a tree of spans: the top-level graph, each model call, and each tool call. For a typical agent that looks like LangGraph → model → ChatAnthropic → tools → search_memory. Every trace carries its status, timestamp, and the agent, workspace, and environment it ran in.

Reading a trace

Open a trace to see the whole run. The header shows its status, duration, span count, tokens, and cost; below it, a waterfall lays the spans out on a timeline so you can see where the time went, and each span shows its own input and output.

Filtering traces

Pick a From and To date to set the window, then narrow the list with Add filter:

• Agent ID, Status, Session ID, User ID, Environment, and Source — exact matches on a run's attributes.
• Workspace ID — scope to a single workspace.
• Feedback — keep only runs an end user rated up or down.
• Full-text and Tool name — search within the run's content.

For anything else, Add key:value matches a metadata key/value pair or a tag. Filters combine with AND; press Search to apply them.