Microsoft Agent Framework Workflows - Human-in-the-loop (HITL)

Let's build a workflow that asks a human operator to guess a number and uses an executor to judge whether the guess is correct.

Enable Request and Response Handling in a Workflow

Requests and responses are handled via a special type called RequestPort.

A RequestPort is a communication channel that allows executors to send requests and receive responses. When an executor sends a message to a RequestPort, the request port emits a RequestInfoEvent that contains the details of the request. External systems can listen for these events, process the requests, and send responses back to the workflow. The framework automatically routes the responses back to the appropriate executor based on the original request.

// Create a request port that receives requests of type NumberSignal and responses of type int.
var numberRequestPort = RequestPort.Create<NumberSignal, int>("GuessNumber");

Add the input port to a workflow.

JudgeExecutor judgeExecutor = new(42);
var workflow = new WorkflowBuilder(numberRequestPort)
    .AddEdge(numberRequestPort, judgeExecutor)
    .AddEdge(judgeExecutor, numberRequestPort)
    .WithOutputFrom(judgeExecutor)
    .Build();

The definition of JudgeExecutor needs a target number and be able to judge whether the guess is correct. If it is not correct, it will send another request to ask for a new guess through the RequestPort.

internal enum NumberSignal
{
    Init,
    Above,
    Below,
}

internal sealed class JudgeExecutor() : Executor<int>("Judge")
{
    private readonly int _targetNumber;
    private int _tries;

    public JudgeExecutor(int targetNumber) : this()
    {
        this._targetNumber = targetNumber;
    }

    public override async ValueTask HandleAsync(int message, IWorkflowContext context, CancellationToken cancellationToken = default)
    {
        this._tries++;
        if (message == this._targetNumber)
        {
            await context.YieldOutputAsync($"{this._targetNumber} found in {this._tries} tries!", cancellationToken);
        }
        else if (message < this._targetNumber)
        {
            await context.SendMessageAsync(NumberSignal.Below, cancellationToken: cancellationToken);
        }
        else
        {
            await context.SendMessageAsync(NumberSignal.Above, cancellationToken: cancellationToken);
        }
    }
}

In Python, executors send requests using ctx.request_info() and handle responses with the @response_handler decorator.

Let's build a workflow that asks a human operator to guess a number and uses an executor to judge whether the guess is correct.

Enable Request and Response Handling in a Workflow

from dataclasses import dataclass

from agent_framework import (
    Executor,
    WorkflowBuilder,
    WorkflowContext,
    handler,
    response_handler,
)


@dataclass
class NumberSignal:
    hint: str  # "init", "above", or "below"


class JudgeExecutor(Executor):
    def __init__(self, target_number: int):
        super().__init__(id="judge")
        self._target_number = target_number
        self._tries = 0

    @handler
    async def handle_guess(self, guess: int, ctx: WorkflowContext[int, str]) -> None:
        self._tries += 1
        if guess == self._target_number:
            await ctx.yield_output(f"{self._target_number} found in {self._tries} tries!")
        elif guess < self._target_number:
            await ctx.request_info(request_data=NumberSignal(hint="below"), response_type=int)
        else:
            await ctx.request_info(request_data=NumberSignal(hint="above"), response_type=int)

    @response_handler
    async def on_human_response(
        self,
        original_request: NumberSignal,
        response: int,
        ctx: WorkflowContext[int, str],
    ) -> None:
        await self.handle_guess(response, ctx)


judge = JudgeExecutor(target_number=42)
workflow = WorkflowBuilder(start_executor=judge).build()

The @response_handler decorator automatically registers the method to handle responses for the specified request and response types. The framework matches incoming responses to the correct handler based on the type annotations of the original_request and response parameters.

An RequestPort emits a RequestInfoEvent when it receives a request. You can subscribe to these events to handle incoming requests from the workflow. When you receive a response from an external system, send it back to the workflow using the response mechanism. The framework automatically routes the response to the executor that sent the original request.

await using StreamingRun handle = await InProcessExecution.RunStreamingAsync(workflow, NumberSignal.Init);
await foreach (WorkflowEvent evt in handle.WatchStreamAsync())
{
    switch (evt)
    {
        case RequestInfoEvent requestInputEvt:
            // Handle `RequestInfoEvent` from the workflow
            int guess = ...; // Get the guess from the human operator or any external system
            await handle.SendResponseAsync(requestInputEvt.Request.CreateResponse(guess));
            break;

        case WorkflowOutputEvent outputEvt:
            // The workflow has yielded output
            Console.WriteLine($"Workflow completed with result: {outputEvt.Data}");
            return;
    }
}

Executors can send requests directly without needing a separate component. When an executor calls ctx.request_info(), the workflow emits a WorkflowEvent with type == "request_info". You can subscribe to these events to handle incoming requests from the workflow. When you receive a response from an external system, send it back to the workflow using the response mechanism. The framework automatically routes the response to the executor's @response_handler method.

from collections.abc import AsyncIterable

from agent_framework import WorkflowEvent


async def process_event_stream(stream: AsyncIterable[WorkflowEvent]) -> dict[str, int] | None:
    """Process events from the workflow stream to capture requests."""
    requests: list[tuple[str, NumberSignal]] = []
    async for event in stream:
        if event.type == "request_info":
            requests.append((event.request_id, event.data))

    # Handle any pending human feedback requests.
    if requests:
        responses: dict[str, int] = {}
        for request_id, request in requests:
            guess = ...  # Get the guess from the human operator or any external system.
            responses[request_id] = guess
        return responses

    return None

# Initiate the first run of the workflow with an initial guess.
# Runs are not isolated; state is preserved across multiple calls to run.
stream = workflow.run(25, stream=True)

pending_responses = await process_event_stream(stream)
while pending_responses is not None:
    # Run the workflow until there is no more human feedback to provide,
    # in which case this workflow completes.
    stream = workflow.run(stream=True, responses=pending_responses)
    pending_responses = await process_event_stream(stream)