agent-framework/dotnet/samples/DurableAgents/ConsoleApps/09_Workflow_Concurrency

Fan-out/Fan-in Workflow with AI Agents

This sample demonstrates the Fan-out/Fan-in pattern using real AI agents in a durable workflow. A question is sent to multiple AI "expert" agents in parallel, and their responses are aggregated into a final result.

Overview

The sample implements an expert consultation workflow using Azure OpenAI:

• A question is parsed and prepared
• The question is sent to 2 AI agents in parallel (Fan-out)
• AI agent responses are collected and aggregated (Fan-in)

Components

Component	Type	Description
ParseQuestion	Executor	Validates and formats the incoming question
Physicist	AI Agent	Azure OpenAI agent with physics expertise
Chemist	AI Agent	Azure OpenAI agent with chemistry expertise
Aggregator	Executor	Combines all AI agent responses

Workflow Pattern

                    ???????????????????
                    ?  ParseQuestion  ?
                    ???????????????????
                             ?
              ???????????????????????????????
              ?                             ?
              ?                             ?
     ???????????????????           ???????????????????
     ?    Physicist    ?           ?     Chemist     ?
     ?   (AI Agent)    ?           ?   (AI Agent)    ?
     ???????????????????           ???????????????????
              ?                             ?
              ?    PARALLEL EXECUTION       ?
              ?                             ?
              ???????????????????????????????
                             ?
                             ?
                  ???????????????????
                  ?   Aggregator    ?
                  ???????????????????

Key Concepts Demonstrated

• Fan-out (AddFanOutEdge): One executor sends to multiple AI agents in parallel
• Fan-in (AddFanInEdge): Multiple AI agent results are collected into an array
• AI Agents in Workflows: Using Azure OpenAI agents as workflow executors
• Durability: If interrupted, completed AI agent responses are preserved on restart

Configuration

When using AI agents in durable workflows, ConfigureDurableWorkflows automatically registers any AI agents found in the workflow. You only need a single configuration call:

services.ConfigureDurableWorkflows(
    options => options.Workflows.AddWorkflow(expertReview),
    workerBuilder: builder => builder.UseDurableTaskScheduler(dtsConnectionString),
    clientBuilder: builder => builder.UseDurableTaskScheduler(dtsConnectionString));

This is similar to how ConfigureDurableOptions works in Azure Functions samples.

Environment Setup

This sample requires:

1. Durable Task Scheduler - See the parent README for setup instructions
2. Azure OpenAI - You need an Azure OpenAI resource with a deployed model

Environment Variables

Variable	Description
DURABLE_TASK_SCHEDULER_CONNECTION_STRING	Connection string for DTS (defaults to local emulator)
AZURE_OPENAI_ENDPOINT	Your Azure OpenAI endpoint URL
AZURE_OPENAI_DEPLOYMENT	Name of your deployed model (e.g., gpt-4)
AZURE_OPENAI_KEY	(Optional) API key - if not set, uses Azure CLI credential

Running the Sample

cd dotnet/samples/DurableAgents/ConsoleApps/09_Workflow_Concurrency
dotnet run --framework net10.0

Sample Session

??????????????????????????????????????????????????????????????????????
?           Fan-out/Fan-in Workflow with AI Agents                   ?
??????????????????????????????????????????????????????????????????????
?  This sample demonstrates parallel AI agent consultation.          ?
?                                                                    ?
?  Workflow: Question ? [Physicist, Chemist] ? Aggregator            ?
?                  (AI agents run in parallel)                       ?
??????????????????????????????????????????????????????????????????????

?? TIP: Stop during execution to test durability - completed agent responses are preserved!

? Checking for pending workflows...

Enter a science question (or 'exit' to quit):
  Example: "What is water?"

Question: What is water

Starting expert review workflow...
Instance ID: abc123-def456-...

???????????????????????????????????????????????????????????????????
? [ParseQuestion] Preparing question for AI agents...
? [ParseQuestion] Question: "What is water?"
? [ParseQuestion] ? Sending to Physicist and Chemist in PARALLEL...
???????????????????????????????????????????????????????????????????

... (AI agents process in parallel)

???????????????????????????????????????????????????????????????????
? [Aggregator] ?? Received 2 AI agent responses
? [Aggregator] Combining into comprehensive answer...
? [Aggregator] ? Aggregation complete!
???????????????????????????????????????????????????????????????????

??????????????????????????????????????????????????????????????????????
?  ? Expert review completed!                                        ?
??????????????????????????????????????????????????????????????????????

???????????????????????????????????????????????????????????????
                    AI EXPERT PANEL RESPONSES
???????????????????????????????????????????????????????????????

?? PHYSICIST:
Water is H2O - two hydrogen atoms bonded to one oxygen atom. From a physics 
perspective, water exhibits unique properties like high specific heat capacity 
and surface tension due to hydrogen bonding.

?? CHEMIST:
Water (H2O) is a polar molecule formed by covalent bonds. Its bent molecular 
geometry creates a dipole moment, making it an excellent solvent for ionic 
and polar compounds.

???????????????????????????????????????????????????????????????
Summary: Received perspectives from 2 AI experts.
???????????????????????????????????????????????????????????????

Durability with AI Agents

If you stop the application while one AI agent is processing:

• The completed agent's response is preserved (cached by Durable Task)
• On restart, only the incomplete agent re-runs
• The aggregator waits for all agents to complete

Related Samples

• 08_SingleWorkflow - Sequential workflow with durability demonstration
• 10_WorkflowConcurrent - Azure Functions version
• 02_AgentOrchestration_Chaining - Agent chaining with durable orchestration