Measure

Agent trace

You might say: span tree, model trace, tool trace

A trace links one run's model requests, tool calls, and hooks into a parent-child timeline. It keeps the agent's internal path attached to the run.

What it does
Shows where the model waited, which tool acted, and which child operation caused a failure or cost.
Ask
Which model call or tool span changed the run's outcome?
Typical
Record a root interaction span with child model, tool, and hook spans. Keep prompts, tool arguments, and results opt-in and redacted by default.
Trouble
Spans start new traces, tool calls are missing, retries are invisible, or sensitive content is recorded without a policy.

What to look for

  1. A root interaction span with nested model and tool spans
  2. Span duration, status, model, tool name, and run identifier
  3. A trace link from the run page to the telemetry backend

Connections

Sources

  1. Anthropic, Monitoring, Span hierarchyEach user prompt starts a `claude_code.interaction` root span. API calls, tool calls, and hook executions are recorded as its children.
  2. Langfuse, Observability data model, Observations and TracesA `trace` represents a single request or operation, for example one chatbot interaction from the user's question to the final response.
  3. Matt Pocock, AI Coding Dictionary, Tool callThe model's output naming a tool and its arguments — just structured text. The harness has to read it and execute.
  4. Matt Pocock, AI Coding Dictionary, Model provider requestOne round-trip from the harness to the model provider. The harness sends the current context; the provider returns one response (a tool call or a final answer).

Use it on your project

teardown-apply: hold "Agent trace" (agent-trace) from the teardown at https://teardownatlas.com/software-factory/ against my project

Paste it into a coding agent that has the teardown skills.

An AI agent wrote this map with the teardown skill. It can contain errors: check the sources before you rely on a part. About and legal