Microsoft Agent Framework workflows now survive process restarts via the Durable Task stack

Shyju Krishnankutty published Durable Workflows in Microsoft Agent Framework on the .NET Blog on May 6, 2026. The headline is that the workflow programming model from Microsoft.Agents.AI.Workflows now plugs into the Durable Task stack via the prerelease Microsoft.Agents.AI.DurableTask package, so a multi-step agent run can checkpoint after every executor and resume on a different process after a crash, scale event, or redeploy. This is the missing piece for taking a multi-agent pipeline from a console demo to something you can actually host.

The piece that was already there

A workflow is a directed graph of Executor<TInput, TOutput> nodes wired with WorkflowBuilder. This part is in the stable Microsoft.Agents.AI.Workflows package and is not new:

using Microsoft.Agents.AI.Workflows;

Workflow cancelOrder = new WorkflowBuilder(orderLookup)
    .WithName("CancelOrder")
    .AddEdge(orderLookup, orderCancel)
    .AddEdge(orderCancel, sendEmail)
    .Build();

await foreach (var evt in InProcessExecution.RunStreamingAsync(cancelOrder, orderId))
{
    Console.WriteLine(evt);
}

InProcessExecution.RunStreamingAsync runs the graph in memory. If the host dies between orderCancel and sendEmail, the order is cancelled, the customer never gets the email, and there is no retry record. Fine for samples, dangerous for anything that touches money.

What changed on May 6

Microsoft.Agents.AI.DurableTask lets the same workflow execute on top of Durable Task. Every executor invocation becomes a Durable activity, so progress is checkpointed after each node and the runtime replays history on restart. From the announcement: “Stateful, durable execution: workflows survive process restarts and failures” and “automatic checkpointing: progress is saved after each step”.

Combined with Microsoft.Agents.AI.Hosting.AzureFunctions and the Microsoft.DurableTask.Client.AzureManaged / Microsoft.DurableTask.Worker.AzureManaged packages targeting Durable Task Scheduler, an Azure Functions host generates the HTTP trigger, orchestrator, and activity functions for you:

var builder = FunctionsApplication.CreateBuilder(args);

builder.ConfigureDurableWorkflows(workflows =>
{
    workflows.AddWorkflow("CancelOrder", sp => BuildCancelOrderWorkflow(sp),
        exposeMcpToolTrigger: true);
});

builder.Build().Run();

exposeMcpToolTrigger: true is the bit worth flagging on its own: the workflow shows up as an MCP tool, so a Claude or Copilot agent can call CancelOrder and the durable runtime handles the rest.

Patterns that get more useful when durability is on

Three workflow features that were nice-to-have in-process become load-bearing once the run can span hours:

• AddFanOutEdge() and AddFanInBarrierEdge() for parallel sub-tasks (e.g. enrich an order from three systems, then merge), now safely resumable.
• RequestPort.Create<TRequest, TResponse>() for human-in-the-loop. The workflow parks, persists, and only wakes when the response arrives. Hours, days, whatever.
• AddSwitch() for conditional routing on a previous executor’s output, which is a much safer pattern than letting the LLM pick the next branch.

The Microsoft.Agents.AI.DurableTask package is still prerelease, so pin the version explicitly and watch the Microsoft Agent Framework DevBlogs feed for breaking changes before GA.