Skip to content

Traces

A trace records what an agent did in one invocation — in post-training, one rollout. It’s a tree of spans: the model calls, tool calls, thoughts, and turns of the agent’s loop, with the prompts, completions, inputs, and outputs each one carried. Traces are logged with the tracing SDK and shown next to the run’s metrics, so a reward curve can be opened into the rollouts that produced it.

A run’s traces are on its Traces tab (8). The tab appears once the run has logged its first trace; while the run is active, the portal checks for one every 15 seconds.

The tab is split in two:

  • Left — a table of the run’s traces or sessions, switched with the Traces | Sessions toggle above it.
  • Right — the trace you picked, read top to bottom as a transcript of its spans. Until you pick one it shows No trace selected.

Drag the divider between them to resize; the portal remembers the size. The open trace, the span you’re on, and the table’s search, filters, and sort are kept in the URL, so you can share a link to exactly what you’re looking at.

Trace : One agent invocation, recorded as the tree under a root span. Each trace is one row in the traces table, belongs to the run, and carries the step — the training step — it was logged at.

Span : One node in a trace. The portal labels each span by its kind:

Portal label What it is SDK kind
LLM call A call to a model — its prompt and completion LLM_CALL
Tool call A tool invocation — its input and output TOOL_CALL
Thought The agent’s reasoning THOUGHT
Turn One pass of the agent’s perceive-act loop TURN
Sub-agent A nested agent, with spans of its own SUB_AGENT
Artifact Something the agent produced ARTIFACT
Function A plain function call FUNCTION
Session summary The record a session leaves when it ends, carrying its reward SESSION_SUMMARY

Status : How a trace ended: Running, Completed, Error, Truncated, or Cancelled.

Session : A conversation with an agent: several traces that share a session ID (the session/id attribute), one per task, plus a summary written when the session ends. The newest summary carries the session’s reward, agent, and number of tasks.

Task : One prompt within a session and the agent’s answer to it, recorded as a trace of its own.