Multi-Agent Session Tracing

Enterprise AI workflows increasingly involve multiple agents collaborating to complete a task. An orchestrator delegates to specialists, validators review outputs, routers fan decisions across services. VeriProof traces the full workflow end to end — even when agents run on separate systems.

How It Works

Session-Level Linking

When an orchestrator agent delegates work to a specialist, VeriProof captures two linked sessions:

The parent session contains an agent_delegation step recording the handoff
The child session carries parentTraceId and parentDelegationSpanId fields linking it back to the delegation step

This linking works across service boundaries. If your orchestrator runs on Service A and the specialist runs on Service B, both sessions appear in the same trace view.

Step-Level Agent Hierarchy

Each step within a session carries agent metadata:

Field	Description
`agentId`	Unique identifier for the agent executing this step
`agentRole`	The agent’s role in the workflow (see below)
`parentStepNumber`	Parent step in the agent hierarchy
`depth`	Nesting level in the multi-agent tree

Agent Role Types

VeriProof recognizes nine agent role types:

Role	Description
`primary`	The main agent handling the user request
`orchestrator`	Coordinates and delegates work to other agents
`specialist`	Performs domain-specific tasks delegated by an orchestrator
`validator`	Reviews and validates outputs from other agents
`reviewer`	Performs quality review of agent work
`router`	Routes requests to appropriate specialist agents
`summarizer`	Aggregates and summarizes outputs from multiple agents
`safety_checker`	Evaluates content safety and policy compliance
`tool_executor`	Executes external tool calls on behalf of other agents

Portal Visualization

Agent Hierarchy Tree

The Time Machine view includes an interactive tree showing parent-child step relationships across agents. Features:

Expand/collapse agent subtrees
Loop abstraction collapses repetitive step cycles into labeled phases
Governance concerns are highlighted per step with visual indicators
Click any node to jump to step detail

Swimlane Diagram

A Gantt-style parallel visualization where each agent gets its own horizontal lane on a shared time axis. This view makes it easy to see:

Which agents ran in parallel vs. sequentially
How long each agent’s work took
Where governance signals fired in the timeline

A sidebar card listing every agent that participated in a session:

Agent ID and role
Number of steps executed
Phase assignments
Governance signal summary per agent

Sub-Agent Session Banner

When viewing a child session created by delegation, the Infrastructure Trace Panel displays a “SUB-AGENT SESSION” banner with context about the parent orchestrator session, including a link to navigate up to the parent.

SDK Integration

Multi-agent tracing is automatic when using VeriProof’s framework-specific adapters:

.NET — AutoGen


// AutoGen multi-agent workflows are automatically instrumented
// See SDK Reference: /sdk/dotnet/autogen

.NET — Semantic Kernel


// Semantic Kernel agent orchestration is traced via the SK adapter
// See SDK Reference: /sdk/dotnet/semantic-kernel

Python — LangGraph


# LangGraph workflows with multiple agents are traced automatically
# See Examples: /examples/python/langgraph

Governance Coverage Per Agent

VeriProof computes governance coverage metrics per agent role across your portfolio. This surfaces gaps such as:

Specialist agents operating without guardrails
Validator agents with no content safety checks
Tool executors lacking grounding verification

Coverage tiers (Foundational, Comprehensive, Advanced) apply individually per agent, giving you a precise view of where governance instrumentation needs attention.

Time Machine Guide — Full session replay documentation
Governance Scoring — How coverage metrics are computed
AutoGen Multi-Agent Example — Complete .NET example
Multi-Turn Workflow — Python multi-agent example