diff --git a/cmd/verifyexamples/examples.go b/cmd/verifyexamples/examples.go
index 01f97b86..d7042d6f 100644
--- a/cmd/verifyexamples/examples.go
+++ b/cmd/verifyexamples/examples.go
@@ -817,6 +817,14 @@ var workflowExamples = []ExampleDefinition{
 		ProjectPath: "examples/03-workflows/loop",
 		MustContain: []string{"found in"},
 	},
+	{
+		Name:        "03_workflows_cancellation",
+		ProjectPath: "examples/03-workflows/cancellation",
+		MustContain: []string{
+			"Cancelling run after",
+			"final run status:",
+		},
+	},
 	{
 		Name:            "03_workflows_shared_states",
 		ProjectPath:     "examples/03-workflows/shared-states",
diff --git a/examples/03-workflows/cancellation/main.go b/examples/03-workflows/cancellation/main.go
new file mode 100644
index 00000000..3c4677d6
--- /dev/null
+++ b/examples/03-workflows/cancellation/main.go
@@ -0,0 +1,164 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"reflect"
+	"time"
+
+	"github.com/microsoft/agent-framework-go/examples/internal/demo"
+	"github.com/microsoft/agent-framework-go/workflow"
+	"github.com/microsoft/agent-framework-go/workflow/inproc"
+)
+
+var _ = demo.NewLogger(
+	"Workflow Cancellation",
+	"This sample streams a long-running workflow, cancels it mid-flight, and inspects the run status.",
+)
+
+// tick carries the current iteration number around the Counter -> Printer loop.
+type tick int
+
+const (
+	// cancelAfter is the number of streamed outputs to observe before cancelling.
+	cancelAfter = 3
+	// workBound is the number of ticks the workflow would emit if left alone.
+	workBound = 100
+	// workDelay simulates a slow, long-running step so the cancellation lands
+	// while the workflow is still producing outputs rather than after it has
+	// already finished. It keeps the sample fully offline.
+	workDelay = 25 * time.Millisecond
+)
+
+func main() {
+	ctx := context.Background()
+
+	// A two-executor loop (Counter -> Printer -> Counter) that would emit
+	// workBound outputs if left to run to completion. We cancel it well before
+	// it reaches the bound.
+	counter := newCounterExecutor("Counter", workBound)
+	printer := newPrinterExecutor("Printer")
+
+	wf, err := workflow.NewBuilder(counter).
+		AddEdge(counter, printer).
+		AddEdge(printer, counter).
+		WithOutputFrom(printer).
+		Build()
+	if err != nil {
+		demo.Panic(err)
+	}
+
+	run, err := inproc.Default.RunStreaming(ctx, wf, tick(1))
+	if err != nil {
+		demo.Panic(err)
+	}
+	defer func() { _ = run.Close(ctx) }()
+
+	var (
+		seen      int
+		cancelled bool
+	)
+stream:
+	for evt, err := range run.WatchStream(ctx) {
+		if err != nil {
+			// Once cancellation is requested, the in-flight superstep unwinds
+			// and surfaces context.Canceled. That is the expected, clean end of
+			// a cancelled stream, so stop iterating rather than treating it as a
+			// failure.
+			if cancelled && errors.Is(err, context.Canceled) {
+				break
+			}
+			demo.Panic(err)
+		}
+		switch e := evt.(type) {
+		case workflow.OutputEvent:
+			seen++
+			demo.Assistantf("output %d: %v", seen, e.Output)
+			if seen == cancelAfter {
+				demo.Assistantf("Cancelling run after %d outputs", seen)
+				if err := run.CancelRun(); err != nil {
+					demo.Panic(err)
+				}
+				cancelled = true
+			}
+		case workflow.ErrorEvent:
+			// A cancelled run reports context.Canceled as it tears down; treat
+			// that as the normal terminal event instead of a hard error.
+			if cancelled && errors.Is(e.Error, context.Canceled) {
+				break stream
+			}
+			demo.Panic(e.Error)
+		case workflow.ExecutorFailedEvent:
+			demo.Panicf("executor %q failed: %v", e.ExecutorID, e.Error)
+		}
+	}
+
+	// WatchStream has returned, so the run has unwound cleanly. The status is
+	// informational only: it is published from a deferred setStatus on the
+	// unwinding goroutine and may still be settling when we read it, so treat
+	// the value as illustrative rather than something to branch on.
+	status, err := run.GetStatus(ctx)
+	if err != nil {
+		demo.Panic(err)
+	}
+	demo.Assistantf("Stream terminated cleanly; final run status: %s", statusLabel(status))
+}
+
+// newCounterExecutor paces the loop: it simulates slow work, then forwards the
+// current tick to the printer until the bound is reached.
+func newCounterExecutor(id string, bound int) workflow.ExecutorBinding {
+	return workflow.NewExecutor(id, func(ctx *workflow.Context, n tick) error {
+		if int(n) > bound {
+			return nil
+		}
+		// Simulate a slow, long-running step, but honor the executor context so a
+		// cancellation lands immediately instead of after the delay elapses.
+		select {
+		case <-time.After(workDelay):
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+		return ctx.SendMessage("", n)
+	}).Extend(&workflow.Executor{
+		ConfigureProtocol: func(rb *workflow.ProtocolBuilder) (*workflow.ProtocolBuilder, error) {
+			rb.SendsMessageType(reflect.TypeFor[tick]())
+			return rb, nil
+		},
+	}).Bind()
+}
+
+// newPrinterExecutor emits one workflow output per tick, then advances the loop.
+func newPrinterExecutor(id string) workflow.ExecutorBinding {
+	return workflow.NewExecutor(id, func(ctx *workflow.Context, n tick) error {
+		if err := ctx.YieldOutput(fmt.Sprintf("tick %d", int(n))); err != nil {
+			return err
+		}
+		return ctx.SendMessage("", n+1)
+	}).Extend(&workflow.Executor{
+		ConfigureProtocol: func(rb *workflow.ProtocolBuilder) (*workflow.ProtocolBuilder, error) {
+			rb.SendsMessageType(reflect.TypeFor[tick]())
+			rb.YieldsOutputType(reflect.TypeFor[string]())
+			return rb, nil
+		},
+	}).Bind()
+}
+
+func statusLabel(status inproc.RunStatus) string {
+	switch status {
+	case inproc.RunStatusNotStarted:
+		return "not started"
+	case inproc.RunStatusIdle:
+		return "idle"
+	case inproc.RunStatusPendingRequests:
+		return "pending requests"
+	case inproc.RunStatusEnded:
+		return "ended"
+	case inproc.RunStatusRunning:
+		return "running"
+	default:
+		return fmt.Sprintf("unknown (%d)", int(status))
+	}
+}