Skip to content

Traces

A trace is the record of one run of your agent or app: each model call, tool call, handoff and sub-step is a span, with its timing, input, output, tokens and errors. Traces come in automatically from the Anosys SDKs, the coding-tool hooks and any OpenTelemetry source.

View traces

Add either widget from Visualizations to a dashboard and point it at your endpoint:

Widget Shows Use it to
Trace Timeline Spans on a time axis Find the slow step
Trace Tree How agents, tools and model calls nest Follow handoffs, subagents and retries

Click any span to see its details: model, input and output, tokens, duration, status and error.

%%{init: {"flowchart": {"curve": "linear"}}}%%
graph TD
    A["Agent run"] --> B["Model call<br/>1.2 s, 1,180 tokens"]
    A --> C["Tool: search<br/>0.8 s, error"]
    A --> D["Handoff to refund agent"]
    D --> E["Model call"]
    D --> F["Tool: issue_refund"]

Trace Tree: a multi-agent run with its agents, turns, model calls and tool calls.

Trace Tree of a multi-agent run with agents, model calls and tool calls

Trace Timeline: the same run on a time axis, with each span's type and duration.

Trace Timeline of a multi-agent run with agent, LLM, handoff and function spans

Find the cause

  • RCA Report widget: a written summary of what likely went wrong in a set of runs.
  • AI Assistant: ask "Why did this run fail?" or "What do the slowest runs have in common?"
  • Contribution Analysis widget: which dimension (model, tool, user, version) contributes most to a change.

Get better traces

  • Tie runs to users. Pass get_user_context in the SDK so you can filter traces by user or session.
  • Name your services. With OpenTelemetry, give each service its own OTEL_SERVICE_NAME.
  • Keep content private if you need to. Turn content capture off in the SDK, or encrypt fields. See Security and data controls.