Multi-agent session tracing
Use linked sessions and step metadata to follow work delegated between agents. The trace can connect activity across services when the integrations report the required relationships. It cannot show unrecorded actions.
Link delegated sessions
A parent session records the handoff with an agent_delegation step.
The child session uses parentTraceId and parentDelegationSpanId to link back to that handoff.
Check both sides of the link when work crosses a service boundary. A matching service name or nearby timestamp does not establish a parent-child relationship.
Describe the agent’s part in the work
Recorded steps can carry these fields:
| Field | Purpose |
|---|---|
agentId | Identifies the agent responsible for the step. |
agentRole | Describes the agent’s function in the workflow. |
parentStepNumber | Links the step to its parent. |
depth | Records its level in the hierarchy. |
Use supported values from the Vocabulary browser. An agent role describes a function; it does not grant portal permissions or prove that a control ran.
Inspect the session in Time Machine
Open a session and choose an available view that explains the work:
- Hierarchy helps you follow parent and delegated steps.
- Swimlane helps compare recorded activity across actors over time.
- Journey presents connected steps and handoffs.
Open the relevant step to inspect its evidence. Follow a linked parent or child session only when your access permits it. See Time Machine for the available views.
Configure each integration
Use the core SDK for each application-owned handoff. Add an Enforce adapter only when the handoff passes through its documented framework boundary. Confirm which operations the integration governs. A linked record does not prove that every action was controlled.
- SDK categories and framework coverage
- SDK package catalog
- TypeScript SDK
- Python SDK
- .NET SDK
- Java SDK
- Rust SDK
- Go SDK
Test a representative handoff. Confirm that both sessions arrived and that their recorded relationship points to the intended delegation.
Review governance evidence per step
Check the policy result, guardrail activity, and review evidence attached to the step under investigation. Do not infer coverage from the agent’s name or role.
For example, naming an agent a validator does not prove that it checked an output. Find the recorded check and its result. If the evidence is missing, ask the application owner whether it was performed and reported.
When totals differ from the expected workflow, inspect missing sessions, broken links, and capture settings before drawing a conclusion about agent behavior.