gabrielmoreira/agent-skills-mirror/agui-dotnet-human-in-the-loop

AG-UI .NET — human in the loop (approval & interrupts)

Goal: stop an agent run at a decision point, hand control to a person, and resume the same run with their decision — either approving a sensitive tool call or supplying input the agent asked for.

A paused run completes its turn with RUN_FINISHED { outcome: interrupt }. The client inspects the response, gets the human decision, then sends a follow-up turn that carries the answer; the SDK reconnects it to the paused tool call so the run continues.

Approval: gate a tool behind a human yes/no

This is the simple path and needs no custom endpoint code — ToChatRequestContext handles the resume.

Server: wrap the sensitive tool

Wrap the function in ApprovalRequiredAIFunction. When the model calls it, the function-invoking client raises a ToolApprovalRequestContent instead of executing:

using Microsoft.Extensions.AI;
 
var deleteFile = new ApprovalRequiredAIFunction(
    AIFunctionFactory.Create(DeleteFile, "delete_file", "Deletes a file."));
 
builder.Services.AddChatClient(/* provider IChatClient */)
    .ConfigureOptions(o => (o.Tools ??= []).Add(deleteFile))
    .UseFunctionInvocation();

Client: detect the request, decide, resume

The first turn returns a ToolApprovalRequestContent. Show it to the human, then resume by appending the request and the human's response and streaming again:

using Microsoft.Extensions.AI;
 
var messages = new List<ChatMessage> { new(ChatRole.User, "Delete report-draft.txt") };
 
var turn1 = new List<ChatResponseUpdate>();
await foreach (var u in client.GetStreamingResponseAsync(messages))
{
    turn1.Add(u);
}
 
var request = turn1.SelectMany(u => u.Contents)
                   .OfType<ToolApprovalRequestContent>()
                   .FirstOrDefault();
 
if (request is { ToolCall: FunctionCallContent call })
{
    bool approved = AskHuman($"Run {call.Name}?");   // your UI / prompt
 
    messages.Add(new ChatMessage(ChatRole.Assistant, [request]));
    messages.Add(new ChatMessage(ChatRole.User, [request.CreateResponse(approved)]));
 
    await foreach (var u in client.GetStreamingResponseAsync(messages))
    {
        Console.Write(u.Text);   // runs the tool if approved, skips it if not
    }
}

On the resumed turn the SDK re-pairs the approval request and response so the function-invoking client executes (or skips) the underlying call.

Interrupt: ask the user for input mid-run

When the agent needs free-form input (not a yes/no), pause with an InterruptRequestContent and resume with an InterruptResponseContent. The client side mirrors approval:

using AGUI.Abstractions;
 
var interrupt = turn1.SelectMany(u => u.Contents)
                     .OfType<InterruptRequestContent>()
                     .FirstOrDefault();
 
if (interrupt is not null)
{
    var answer = AskHuman(interrupt.Message);   // e.g. a username
    var payload = JsonSerializer.SerializeToElement(new { response = answer });
 
    messages.Add(new ChatMessage(ChatRole.Assistant, [interrupt]));
    messages.Add(new ChatMessage(ChatRole.User,
        [new InterruptResponseContent(interrupt.RequestId) { Payload = payload }]));
 
    await foreach (var u in client.GetStreamingResponseAsync(messages))
    {
        Console.Write(u.Text);
    }
}

AGUIChatClient encodes the response as RunAgentInput.Resume[] on the wire. The interrupt request carries a Reason (e.g. InterruptReasons.InputRequired), a Message to display, and an optional ResponseSchema describing the expected payload.

On the server, raise the interrupt by yielding a content-bearing update from a DelegatingChatClient. Typically you bridge a model tool call into the interrupt: detect the model's FunctionCallContent, emit an InterruptRequestContent(call.CallId) instead, and on resume rewrite the InterruptResponseContent back into the FunctionCallContent + FunctionResultContent pair the model expects, so it continues as if the tool returned the user's answer.

Anti-patterns

• Resuming with only the response content. Both the request (ToolApprovalRequestContent / InterruptRequestContent) and the matching response must be in the resumed message list, with the same id; the SDK pairs them to reconnect to the paused call. Sending the response alone leaves the run with nothing to resume.
• Instructing the model to ask for confirmation in its prompt. The approval gate is structural — the wrapped tool pauses the run on its own. A prompt that also tells the model to "ask the user to confirm" produces a redundant text question and a second round-trip. Tell the model to call the tool directly and let the gate handle approval.

Verify

1. First turn ends without running the side effect: the stream finishes RUN_FINISHED { outcome: interrupt } and the response contains a ToolApprovalRequestContent (or InterruptRequestContent).
2. After resuming with approval, the tool's effect occurs and the run finishes normally; after resuming with a rejection, the effect does not occur and the model proceeds without it.
3. For input interrupts, the agent's final answer reflects the value the human supplied.