MultiAgentRouting

Multi-Agent Routing: Sequential and Parallel Dispatch

Overview

Demonstrates two complementary patterns for dispatching to multiple agents: sequential routing (classify then dispatch to one specialist) and parallel fan-out (send the same query to all agents simultaneously). Both patterns include OpenTelemetry tracing wired through all four Temporal + agent span sources.

This sample demonstrates:

• RoutingWorkflow: keyword-based classification via a RoutingActivities.ClassifyRequest activity, then dispatch to one of three specialists
• ParallelAgentWorkflow: fan-out to all three specialists at once via ExecuteAgentsInParallelAsync
• OTel tracing with TracingInterceptor and TemporalAgentTelemetry span sources
• Routing decision as a durable, auditable activity result

Architecture

User Question
    │
    ▼
RoutingWorkflow                         ParallelAgentWorkflow
    │                                       │
    ├─ Activity: ClassifyRequest()          ├─ GetAgent("WeatherAgent")    ─┐
    │    └─ returns: "WeatherAgent"         ├─ GetAgent("BillingAgent")    ─┼─ ExecuteAgentsInParallelAsync
    │                                       └─ GetAgent("TechSupportAgent") ┘
    └─ GetAgent(agentName).RunAsync()              │
         └─ specialist response               IReadOnlyList<AgentResponse>

OTel Span Hierarchy

agent.client.send          (TemporalAgentTelemetry — agent name, session ID)
  UpdateWorkflow:RunAgent  (TracingInterceptor — workflow update span)
    RunActivity:ExecuteAgent  (TracingInterceptor — activity span)
      agent.turn            (TemporalAgentTelemetry — token counts, correlation ID)

All four sources must be registered for the full trace to appear:

Sdk.CreateTracerProviderBuilder()
    .AddSource(TracingInterceptor.ClientSource.Name)
    .AddSource(TracingInterceptor.WorkflowsSource.Name)
    .AddSource(TracingInterceptor.ActivitiesSource.Name)
    .AddSource(TemporalAgentTelemetry.ActivitySourceName)
    .AddConsoleExporter()
    .Build();

Highlights

• Routing decision as activity. ClassifyRequest runs in a RoutingActivities activity, not inline in the workflow. The result is cached in event history — a crash after classification won't re-invoke the classifier, and the decision is visible in the Temporal Web UI.
• Consistent fallback agent. ClassifyRequest returns TechSupportAgent for both empty input and for input where no keywords match any specialist. This makes TechSupportAgent the single general-purpose fallback in all unresolved cases.
• ExecuteAgentsInParallelAsync for fan-out. The workflow-safe equivalent of Task.WhenAll — dispatches multiple agent activities concurrently and returns results in input order. Each agent runs in its own session; results come back as IReadOnlyList<AgentResponse> in the same order the agents were passed in.
• .ConfigureAwait(true) on all workflow awaits. All await calls inside [Workflow]-attributed classes use .ConfigureAwait(true). This is required to keep the Temporal workflow task scheduler active — omitting it can cause the workflow context to be lost during replay.
• TracingInterceptor propagates context. Registered on ITemporalClient via opts.Interceptors, it propagates OTel trace context across Temporal's RPC boundary so spans from client to workflow to activity form a single trace.
• Three registered specialists. WeatherAgent, BillingAgent, and TechSupportAgent are all added via AddTemporalAgents() on the same worker, demonstrating multi-agent registration.

Getting Started

Prerequisites

• .NET 10 SDK or later
• A local Temporal server: temporal server start-dev
• An OpenAI-compatible API key

Configure API credentials

OPENAI_API_KEY is required and validated first on startup. OPENAI_API_BASE_URL is also required.

dotnet user-secrets set "OPENAI_API_KEY" "sk-..." --project samples/MAF/MultiAgentRouting
dotnet user-secrets set "OPENAI_API_BASE_URL" "https://api.openai.com/v1" --project samples/MAF/MultiAgentRouting

Run

dotnet run --project samples/MAF/MultiAgentRouting/MultiAgentRouting.csproj

Expected Output

Worker started.

── Demonstrating workflow-based routing ────────────────────

User: Will it rain in Seattle tomorrow?
Agent: <WeatherAgent response>

User: Why was I charged twice on my last invoice?
Agent: <BillingAgent response>

User: My application keeps crashing with a null reference exception.
Agent: <TechSupportAgent response>

── Demonstrating parallel agent execution ──────────────────

Fan-out query (sent to all 3 agents simultaneously): "Briefly introduce yourself..."

Parallel responses:

[WeatherAgent]: I'm a weather specialist...
[BillingAgent]: I handle billing and payments...
[TechSupportAgent]: I provide technical support...

Done.