Trace and Conversation Structure

What a trace contains, how to read one, and how traces relate to conversations.

Every agent run is captured as a trace, and every multi-turn exchange as a conversation. They're the evidence behind everything else in Agent Trust: monitors alert on them, Conversation Clusters group them, and Reinforcement Loop cites them in every finding. This page covers how to find, read, and export them when you want to dig in yourself.

Find the trace you need

Start from the Agent summary to see where an agent is degrading — trace count, latency percentiles, token usage, and error rates over a time range, segmentable by workflow, task, or model.

Then drill into the Trace Explorer, which lists every trace with its status, duration, token usage, and error state.

  • Search on your own metadata. Use key:value on any attribute your agent emits, plus built-in fields. Prefix with - to exclude, quote values with spaces, and chain with AND.
  • Filter on errors two ways. Final error state shows runs that ultimately failed. Has errors anywhere also catches runs that hit an error mid-flight and recovered — the ones an "OK" status would hide.
  • Compare side by side. Use the open-in-new-tab icon on a trace row to open several traces without losing your filtered view.

    📘

    Custom-metadata search is available for OTel agents on the Agent Observability Data Platform.

Read a trace

A trace is a tree of spans, each one unit of work. Click any span to see its inputs, outputs, timing, and metadata.

Span typeWhat it captures
LLM spanPrompt, completion, model name, token counts, latency
Tool spanTool name, the inputs the agent passed, the response that came back, duration
Custom spanAny step your agent instruments

Every span also carries timing, status, and any custom attributes you emit, like ticket IDs or customer names.

Start with tool spans. They show what the agent actually did — queried a table, called an API, ran a function — and what it got back, which is usually where the problem is.

Don't want to walk the span tree? Explain this trace gives you a plain-English summary of what the agent did, ready to share.

Read a conversation

Chat-style agents produce one trace per user turn. The conversation view stitches them into a single dialogue, so you can follow the back-and-forth without jumping between traces.

  • Messages appear in order, labeled user, agent, system, or tool. Tool messages include the tool name, inputs, output, and duration.
  • Click any message to jump to the span behind it.
  • Explain this conversation summarizes the whole exchange.
  • Evaluate on demand. Run an evaluation on any conversation — pick from the standard set or write your own prompt — and get results in seconds, no monitor needed. Use it to check "did the agent do the right thing here?" or to test an eval before turning it into an Agent Evaluation Monitor.

Traces are grouped into a conversation by a shared conversation ID. For OpenTelemetry agents, set gen_ai.conversation.id on every span — see OpenTelemetry Integration.

Errors

An agent's status reflects its final outcome. Errors that happened mid-run are tracked separately, so a run that recovered still shows what went wrong. Error counts match across the trace list, agent summary, dashboards, and conversations.

To get alerted when error rates rise, use an Agent Metric Monitor on error rate.

Export traces

Turn real production runs into eval sets, fine-tuning data, or offline repro cases.

  • Selected spans as JSONL. From a trace, select the LLM spans you want and export them — one record per span, with input, output, tool calls, and model content.
  • A full trace. Use Export trace in the UI, or the CLI, to get a gzipped JSON of the whole span tree with per-span content:
montecarlo agent-traces export --trace-link <copy-link-from-UI>

Related


Did this page help you?