Workflows

Overview

Workflows in the Microsoft Agent Framework .NET SDK enable you to build sophisticated multi-step agentic applications. Workflows orchestrate multiple processing units (executors) connected by edges, supporting patterns like sequential processing, parallel execution, conditional routing, and loops.

Core Concepts

Executors

Executors are the processing units in a workflow. Each executor:

Accepts strongly-typed input
Performs processing (calls an agent, transforms data, etc.)
Returns strongly-typed output
Can be stateful or stateless

public abstract class Executor<TIn, TOut>
{
    public abstract ValueTask<TOut> HandleAsync(
        TIn message,
        IWorkflowContext context,
        CancellationToken cancellationToken = default);
}

Edges

Edges define the data flow between executors. They determine:

Which executors are connected
The order of execution
Conditional routing logic

WorkflowBuilder

WorkflowBuilder is the fluent API for constructing workflows:

var builder = new WorkflowBuilder(entryPointExecutor);
builder.AddEdge(executorA, executorB);
builder.WithOutputFrom(finalExecutor);
var workflow = builder.Build();

Creating Executors

Custom Executor Class

using Microsoft.Agents.AI.Workflows;

public class UppercaseExecutor : Executor<string, string>
{
    public UppercaseExecutor() : base("UppercaseExecutor") { }
    
    public override ValueTask<string> HandleAsync(
        string message,
        IWorkflowContext context,
        CancellationToken cancellationToken = default)
    {
        return ValueTask.FromResult(message.ToUpperInvariant());
    }
}

From Lambda Functions

Use BindAsExecutor() to create executors from lambda functions:

Func<string, string> uppercaseFunc = s => s.ToUpperInvariant();
var uppercase = uppercaseFunc.BindAsExecutor("UppercaseExecutor");

Func<string, string> reverseFunc = s => string.Concat(s.Reverse());
var reverse = reverseFunc.BindAsExecutor("ReverseExecutor");

From Async Functions

Func<string, Task<string>> asyncFunc = async s =>
{
    await Task.Delay(100);
    return s.ToUpperInvariant();
};
var executor = asyncFunc.BindAsExecutor("AsyncExecutor");

Building Workflows

Simple Sequential Workflow

using Microsoft.Agents.AI.Workflows;

// Create executors
Func<string, string> uppercaseFunc = s => s.ToUpperInvariant();
var uppercase = uppercaseFunc.BindAsExecutor("UppercaseExecutor");

Func<string, string> reverseFunc = s => string.Concat(s.Reverse());
var reverse = reverseFunc.BindAsExecutor("ReverseExecutor");

// Build workflow: input -> uppercase -> reverse -> output
var builder = new WorkflowBuilder(uppercase);
builder.AddEdge(uppercase, reverse);
builder.WithOutputFrom(reverse);
var workflow = builder.Build();

// Execute workflow
await using Run run = await InProcessExecution.RunAsync(
    workflow, 
    "Hello, World!");

foreach (WorkflowEvent evt in run.NewEvents)
{
    if (evt is ExecutorCompletedEvent completed)
    {
        Console.WriteLine($"{completed.ExecutorId}: {completed.Data}");
    }
}
// Output:
// UppercaseExecutor: HELLO, WORLD!
// ReverseExecutor: !DLROW ,OLLEH

Workflow with Agents

Integrate AI agents into workflows:

using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Workflows;

// Create an agent
AIAgent writerAgent = chatClient.AsAIAgent(
    name: "Writer",
    instructions: "You are a creative writer. Write short stories.");

AIAgent editorAgent = chatClient.AsAIAgent(
    name: "Editor",
    instructions: "You are an editor. Improve the writing quality.");

// Bind agents as executors
var writer = writerAgent.BindAsExecutor("WriterStep");
var editor = editorAgent.BindAsExecutor("EditorStep");

// Build workflow
var builder = new WorkflowBuilder(writer);
builder.AddEdge(writer, editor);
builder.WithOutputFrom(editor);
var workflow = builder.Build();

// Execute
await using var run = await InProcessExecution.RunAsync(
    workflow,
    "Write a short story about a robot learning to paint.");

foreach (var evt in run.NewEvents)
{
    if (evt is AgentResponseEvent agentEvent)
    {
        Console.WriteLine($"{agentEvent.AgentName}: {agentEvent.Response.Text}");
    }
}

Streaming Workflow Events

Stream workflow events as they occur:

await using var run = await InProcessExecution.RunAsync(workflow, input);

await foreach (var evt in run.StreamAsync())
{
    switch (evt)
    {
        case ExecutorStartedEvent started:
            Console.WriteLine($"Started: {started.ExecutorId}");
            break;
        
        case ExecutorCompletedEvent completed:
            Console.WriteLine($"Completed: {completed.ExecutorId}");
            Console.WriteLine($"Result: {completed.Data}");
            break;
        
        case AgentResponseUpdateEvent update:
            Console.Write(update.Update.Text);
            break;
        
        case ExecutorFailedEvent failed:
            Console.WriteLine($"Error: {failed.Exception.Message}");
            break;
    }
}

Parallel Execution (Fan-Out/Fan-In)

Execute multiple executors in parallel:

// Create parallel executors
var executor1 = func1.BindAsExecutor("Executor1");
var executor2 = func2.BindAsExecutor("Executor2");
var executor3 = func3.BindAsExecutor("Executor3");

// Create aggregator to collect results
var aggregator = new AggregatingExecutor<string, string>("Aggregator");

// Build workflow with fan-out pattern
var builder = new WorkflowBuilder(inputExecutor);
builder.AddEdge(inputExecutor, executor1);
builder.AddEdge(inputExecutor, executor2);
builder.AddEdge(inputExecutor, executor3);

// Fan-in: collect results
builder.AddEdge(executor1, aggregator);
builder.AddEdge(executor2, aggregator);
builder.AddEdge(executor3, aggregator);

builder.WithOutputFrom(aggregator);
var workflow = builder.Build();

Conditional Edges

Route based on executor output:

public class RouterExecutor : Executor<int, string>
{
    public RouterExecutor() : base("Router") { }
    
    public override ValueTask<string> HandleAsync(
        int value,
        IWorkflowContext context,
        CancellationToken cancellationToken = default)
    {
        return ValueTask.FromResult(value > 0 ? "positive" : "negative");
    }
}

// Build workflow with conditional routing
var router = new RouterExecutor();
var positiveHandler = positiveFunc.BindAsExecutor("PositiveHandler");
var negativeHandler = negativeFunc.BindAsExecutor("NegativeHandler");

var builder = new WorkflowBuilder(router);

// Add conditional edges
builder.AddEdge(
    router, 
    positiveHandler, 
    condition: output => output == "positive");

builder.AddEdge(
    router, 
    negativeHandler, 
    condition: output => output == "negative");

builder.WithOutputFrom(positiveHandler);
builder.WithOutputFrom(negativeHandler);
var workflow = builder.Build();

Loops

Create iterative workflows:

public class CounterExecutor : Executor<int, int>
{
    public CounterExecutor() : base("Counter") { }
    
    public override ValueTask<int> HandleAsync(
        int count,
        IWorkflowContext context,
        CancellationToken cancellationToken = default)
    {
        return ValueTask.FromResult(count + 1);
    }
}

var counter = new CounterExecutor();
var checker = new CheckerExecutor(); // Checks if count < 5

var builder = new WorkflowBuilder(counter);

// Create loop: counter -> checker -> (back to counter if < 5)
builder.AddEdge(counter, checker);
builder.AddEdge(
    checker, 
    counter, 
    condition: count => count < 5);

builder.WithOutputFrom(checker);
var workflow = builder.Build();

Writer-Critic Pattern

Iterative refinement with quality gates:

// Create agents
AIAgent writerAgent = chatClient.AsAIAgent(
    name: "Writer",
    instructions: "Generate creative content based on feedback.");

AIAgent criticAgent = chatClient.AsAIAgent(
    name: "Critic",
    instructions: "Review content and provide feedback. Rate quality 1-10.");

// Bind as executors
var writer = writerAgent.BindAsExecutor("Writer");
var critic = criticAgent.BindAsExecutor("Critic");

// Quality gate executor
var qualityGate = new QualityGateExecutor(); // Checks if rating >= 8

var builder = new WorkflowBuilder(writer);
builder.AddEdge(writer, critic);
builder.AddEdge(critic, qualityGate);

// Loop back if quality not met (max 3 iterations)
builder.AddEdge(
    qualityGate,
    writer,
    condition: result => result.Quality < 8 && result.Iteration < 3);

builder.WithOutputFrom(qualityGate);
var workflow = builder.Build();

await using var run = await InProcessExecution.RunAsync(
    workflow,
    "Write a haiku about AI");

Checkpointing

Save and restore workflow state:

using Microsoft.Agents.AI.Workflows.Checkpointing;

// Create checkpointer
var checkpointer = new InMemoryCheckpointer();

// Run workflow with checkpointing
await using var run = await InProcessExecution.RunAsync(
    workflow,
    input,
    checkpointer: checkpointer);

// Save checkpoint
var checkpointId = await run.CheckpointAsync();

// Later, restore from checkpoint
await using var restoredRun = await InProcessExecution.ResumeAsync(
    workflow,
    checkpointId,
    checkpointer: checkpointer);

Human-in-the-Loop

Pause workflow for human input:

public class ApprovalExecutor : Executor<string, string>
{
    public ApprovalExecutor() : base("Approval") { }
    
    public override async ValueTask<string> HandleAsync(
        string content,
        IWorkflowContext context,
        CancellationToken cancellationToken = default)
    {
        // Request human approval
        var approvalRequest = new ApprovalRequestEvent
        {
            Content = content,
            RequestId = Guid.NewGuid().ToString()
        };
        
        context.AddEvent(approvalRequest);
        
        // Wait for human response via input port
        var response = await context.WaitForInputAsync<ApprovalResponse>(
            approvalRequest.RequestId,
            cancellationToken);
        
        return response.Approved ? content : "REJECTED";
    }
}

// Build workflow with approval
var builder = new WorkflowBuilder(processorExecutor);
builder.AddEdge(processorExecutor, approvalExecutor);
builder.AddEdge(approvalExecutor, finalExecutor);
builder.WithOutputFrom(finalExecutor);
var workflow = builder.Build();

// Execute workflow
await using var run = await InProcessExecution.RunAsync(workflow, input);

// Monitor for approval requests
await foreach (var evt in run.StreamAsync())
{
    if (evt is ApprovalRequestEvent approvalReq)
    {
        Console.WriteLine($"Approval needed: {approvalReq.Content}");
        Console.Write("Approve? (y/n): ");
        var approved = Console.ReadLine()?.ToLower() == "y";
        
        // Send approval response
        await run.SendInputAsync(
            approvalReq.RequestId,
            new ApprovalResponse { Approved = approved });
    }
}

Shared State

Share state between executors:

public class StateExecutor : Executor<string, string>
{
    public override ValueTask<string> HandleAsync(
        string message,
        IWorkflowContext context,
        CancellationToken cancellationToken = default)
    {
        // Access shared state
        var state = context.GetState<SharedData>("myState") 
            ?? new SharedData();
        
        state.Counter++;
        state.Messages.Add(message);
        
        // Update shared state
        context.SetState("myState", state);
        
        return ValueTask.FromResult($"Processed {state.Counter} messages");
    }
}

public class SharedData
{
    public int Counter { get; set; }
    public List<string> Messages { get; set; } = new();
}

Sub-Workflows

Compose workflows hierarchically:

// Create sub-workflow
var subBuilder = new WorkflowBuilder(subExecutor1);
subBuilder.AddEdge(subExecutor1, subExecutor2);
subBuilder.WithOutputFrom(subExecutor2);
var subWorkflow = subBuilder.Build();

// Bind sub-workflow as executor
var subWorkflowExecutor = subWorkflow.BindAsExecutor("SubWorkflow");

// Use in parent workflow
var parentBuilder = new WorkflowBuilder(parentExecutor1);
parentBuilder.AddEdge(parentExecutor1, subWorkflowExecutor);
parentBuilder.AddEdge(subWorkflowExecutor, parentExecutor2);
parentBuilder.WithOutputFrom(parentExecutor2);
var parentWorkflow = parentBuilder.Build();

Workflow Visualization

Generate workflow diagrams:

using Microsoft.Agents.AI.Workflows.Visualization;

// Generate Mermaid diagram
string mermaid = workflow.ToMermaid();
Console.WriteLine(mermaid);

// Generate DOT graph
string dot = workflow.ToDot();
File.WriteAllText("workflow.dot", dot);

Best Practices

Keep Executors Focused

Each executor should have a single, well-defined responsibility. Break complex logic into multiple executors.

Use Strong Types

Leverage .NET’s type system with Executor<TIn, TOut> to catch errors at compile time:

public class Executor1 : Executor<string, int> { }
public class Executor2 : Executor<int, string> { }

// Type-safe edge
builder.AddEdge(executor1, executor2); // TOut of executor1 matches TIn of executor2

Handle Errors Gracefully

Implement error handling in executors:

public override async ValueTask<string> HandleAsync(
    string input,
    IWorkflowContext context,
    CancellationToken cancellationToken)
{
    try
    {
        return await ProcessAsync(input);
    }
    catch (Exception ex)
    {
        context.AddEvent(new ErrorEvent { Error = ex.Message });
        return "ERROR";
    }
}

Add Checkpoints for Long Workflows

Use checkpointing for workflows that take significant time or process critical data.

Monitor Workflow Events

Subscribe to workflow events for observability:

await foreach (var evt in run.StreamAsync())
{
    Logger.LogInformation(
        "Event: {Type} at {Time}",
        evt.GetType().Name,
        DateTime.UtcNow);
}

Use Conditional Edges for Complex Routing

Prefer conditional edges over complex executor logic for routing decisions.

Limit Loop Iterations

Always include iteration limits in loops to prevent infinite execution:

builder.AddEdge(
    executor,
    loopExecutor,
    condition: state => state.Iteration < MAX_ITERATIONS);

Next Steps

Declarative Workflows

Define workflows using YAML

Durable Workflows

Build reliable, resumable workflows

Multi-Agent Patterns

Common agentic workflow patterns

Observability

Monitor and debug workflows

Getting Started

Guides

API Reference

Overview

Core Concepts

Executors

Edges

WorkflowBuilder

Creating Executors

Custom Executor Class

From Lambda Functions

From Async Functions

Building Workflows

Simple Sequential Workflow

Workflow with Agents

Streaming Workflow Events

Parallel Execution (Fan-Out/Fan-In)

Conditional Edges

Loops

Writer-Critic Pattern

Checkpointing

Human-in-the-Loop

Shared State

Sub-Workflows

Workflow Visualization

Best Practices

Next Steps

Declarative Workflows

Durable Workflows

Multi-Agent Patterns

Observability

Build docs developers (and LLMs) love

Getting Started

Guides

API Reference

​Overview

​Core Concepts

​Executors

​Edges

​WorkflowBuilder

​Creating Executors

​Custom Executor Class

​From Lambda Functions

​From Async Functions

​Building Workflows

​Simple Sequential Workflow

​Workflow with Agents

​Streaming Workflow Events

​Parallel Execution (Fan-Out/Fan-In)

​Conditional Edges

​Loops

​Writer-Critic Pattern

​Checkpointing

​Human-in-the-Loop

​Shared State

​Sub-Workflows

​Workflow Visualization

​Best Practices

​Next Steps

Declarative Workflows

Durable Workflows

Multi-Agent Patterns

Observability

Build docs developers (and LLMs) love

Overview

Core Concepts

Executors

Edges

WorkflowBuilder

Creating Executors

Custom Executor Class

From Lambda Functions

From Async Functions

Building Workflows

Simple Sequential Workflow

Workflow with Agents

Streaming Workflow Events

Parallel Execution (Fan-Out/Fan-In)

Conditional Edges

Loops

Writer-Critic Pattern

Checkpointing

Human-in-the-Loop

Shared State

Sub-Workflows

Workflow Visualization

Best Practices

Next Steps