diff --git a/config/crd/bases/anesthesia.k3s.fr_anesthesianodes.yaml b/config/crd/bases/anesthesia.k3s.fr_anesthesianodes.yaml
index f55f4b4..75bd850 100644
--- a/config/crd/bases/anesthesia.k3s.fr_anesthesianodes.yaml
+++ b/config/crd/bases/anesthesia.k3s.fr_anesthesianodes.yaml
@@ -79,8 +79,9 @@ spec:
                 - unknown
                 - starting
                 - running
+                - draining
                 - shutting-down
-                - "off"
+                - halted
                 type: string
             type: object
         type: object
diff --git a/controllers/anesthesianode_controller.go b/controllers/anesthesianode_controller.go
index f06f5dc..55fa228 100644
--- a/controllers/anesthesianode_controller.go
+++ b/controllers/anesthesianode_controller.go
@@ -52,6 +52,7 @@ var (
 	defaultCidrV4      = "/24"
 	defaultCidrV6      = "/56"
 	wolImage           = "registry.k3s.fr/k8s/docker-wakeonlan-image/main:latest"
+	shutdownImage      = "registry.k3s.fr/k8s/docker-ubuntu-systemd/main:latest"
 	namespace          = "default"
 )
 
@@ -104,6 +105,13 @@ func jobLabels(node *anesthesiav1alpha1.AnesthesiaNode) map[string]string {
 	}
 }
 
+func podLabels(node *anesthesiav1alpha1.AnesthesiaNode) map[string]string {
+	return map[string]string{
+		"anesthesia.k3s.fr/node": node.Name,
+		"anesthesia.k3s.fr/type": "shutdown",
+	}
+}
+
 func (r *AnesthesiaNodeReconciler) WolJob(runnerNode *corev1.Node, wakeupNode *anesthesiav1alpha1.AnesthesiaNode) *batch.Job {
 	randid := uuid.New().String()[0:6]
 
@@ -133,11 +141,66 @@ func (r *AnesthesiaNodeReconciler) WolJob(runnerNode *corev1.Node, wakeupNode *a
 
 }
 
+func shutdownPod(shutdownNode *anesthesiav1alpha1.AnesthesiaNode) *corev1.Pod {
+	randid := uuid.New().String()[0:6]
+
+	hostPathDirectory := corev1.HostPathDirectory
+
+	pod := &corev1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "shutdown-" + shutdownNode.Name + "-" + randid,
+			Namespace: namespace,
+			Labels:    podLabels(shutdownNode),
+		},
+		Spec: corev1.PodSpec{
+			NodeName:      shutdownNode.Name,
+			RestartPolicy: "Never",
+			HostPID:       true,
+			Containers: []corev1.Container{
+				{
+					Image:   shutdownImage,
+					Name:    "shutdown",
+					Command: []string{"sh", "-c", "systemctl poweroff"},
+					Env: []corev1.EnvVar{
+						{
+							Name:  "SYSTEMD_IGNORE_CHROOT",
+							Value: "1",
+						},
+					},
+					VolumeMounts: []corev1.VolumeMount{
+						{
+							MountPath: "/run/systemd",
+							Name:      "runsystemd",
+						},
+					},
+				},
+			},
+			Volumes: []corev1.Volume{
+				{
+					Name: "runsystemd",
+					VolumeSource: corev1.VolumeSource{
+						HostPath: &corev1.HostPathVolumeSource{
+							Path: "/run/systemd",
+							Type: &hostPathDirectory,
+						},
+					},
+				},
+			},
+		},
+	}
+	return pod
+}
+
 // Ensures that there is a WOL pod running on a targetNode (that is on the same subnet)
+// XXX to rebase ensureRessource
 func (r *AnesthesiaNodeReconciler) ensureWolJob(job *batch.Job) error {
 	err := r.Create(context.TODO(), job)
 	return err
 }
+func (r *AnesthesiaNodeReconciler) ensurePod(job *corev1.Pod) error {
+	err := r.Create(context.TODO(), job)
+	return err
+}
 
 // Returns true if both IPs are on the same /cidr subnet
 func areIpsOnSameNet(ip1 string, ip2 string, cidr string) bool {
@@ -206,6 +269,7 @@ func (r *AnesthesiaNodeReconciler) findNodesOnSameL2(n *corev1.Node) ([]*corev1.
 			sameL2Nodes = append(sameL2Nodes, &nn)
 		}
 	}
+	fmt.Printf("SameL2as %v: %v\n\n\n", n, sameL2Nodes)
 	return sameL2Nodes, nil
 }
 
@@ -231,6 +295,7 @@ func (r *AnesthesiaNodeReconciler) listWakeupJobs(wakeupNode *anesthesiav1alpha1
 }
 
 func (r *AnesthesiaNodeReconciler) areRunningWakeupJobs(wakeupNode *anesthesiav1alpha1.AnesthesiaNode) (bool, error) {
+	// XXX this could return the running wakeupjobs so that the function is not called twice
 	nodeJobs, err := r.listWakeupJobs(wakeupNode)
 	if err != nil {
 		return false, err
@@ -267,6 +332,52 @@ func (r *AnesthesiaNodeReconciler) removeOldJobs(wakeupNode *anesthesiav1alpha1.
 	return nil
 }
 
+func (r *AnesthesiaNodeReconciler) listShutdownPods(shutdownNode *anesthesiav1alpha1.AnesthesiaNode) (*corev1.PodList, error) {
+	nodePodLabels := podLabels(shutdownNode)
+	var podList corev1.PodList
+
+	err := r.List(context.TODO(), &podList, client.InNamespace(namespace), client.MatchingLabels(nodePodLabels))
+	return &podList, err
+
+}
+
+func (r *AnesthesiaNodeReconciler) removePod(pod corev1.Pod) error {
+	fmt.Printf("Will remove pod %v\n", pod.Name)
+	err := r.Delete(context.TODO(), &pod)
+	return err
+
+}
+
+func (r *AnesthesiaNodeReconciler) removeOldShutdownPods(shutdownNode *anesthesiav1alpha1.AnesthesiaNode) error {
+	shutdownPods, err := r.listShutdownPods(shutdownNode)
+	if err != nil {
+		fmt.Printf("Could not list shutdown pods: %v\n", err)
+		return err
+	}
+
+	for _, pod := range shutdownPods.Items {
+		switch pod.Status.Phase {
+		case corev1.PodPending, corev1.PodRunning:
+			// Nothing to do
+		case corev1.PodFailed:
+			fmt.Printf("Pod %v failed !\n", pod.Name)
+			fallthrough
+		case corev1.PodUnknown:
+			// Shouldn't be used by k8s, but arises in k3s
+			fallthrough
+		case corev1.PodSucceeded:
+			// This pod shall be removed
+			err := r.removePod(pod)
+			if err != nil {
+				return err // Should we continue with others instead ?
+			}
+		}
+
+	}
+	return nil
+
+}
+
 //+kubebuilder:rbac:groups=anesthesia.k3s.fr,resources=anesthesianodes,verbs=get;list;watch;create;update;patch;delete
 //+kubebuilder:rbac:groups=anesthesia.k3s.fr,resources=anesthesianodes/status,verbs=get;update;patch
 //+kubebuilder:rbac:groups=anesthesia.k3s.fr,resources=anesthesianodes/finalizers,verbs=update
@@ -288,6 +399,7 @@ func (r *AnesthesiaNodeReconciler) Reconcile(ctx context.Context, req ctrl.Reque
 	if req.Name != "think02" {
 		return ctrl.Result{}, nil
 	}
+	fmt.Printf("\n\n\nHALLO, I was launched !!!\n\n")
 
 	if err := r.Get(ctx, req.NamespacedName, &node); err != nil {
 		log.Error(err, "Could not get AnesthesiaNode")
@@ -302,14 +414,14 @@ func (r *AnesthesiaNodeReconciler) Reconcile(ctx context.Context, req ctrl.Reque
 
 	renewTime, err := r.lastHeartbeat(v1node)
 	if err != nil {
-		log.Error(err, "")
+		log.Error(err, "Couldn't get last heartbeat")
 		return ctrl.Result{}, nil
 	}
 
 	nodeAlive := isNodeAlive(renewTime)
 
 	var newActualState string
-	requeue := 2 * time.Second
+	requeue := 0 * time.Second
 	if node.Spec.State == "on" && nodeAlive { // FIXME
 		newActualState = "running"
 		err = r.removeOldJobs(&node)
@@ -334,14 +446,16 @@ func (r *AnesthesiaNodeReconciler) Reconcile(ctx context.Context, req ctrl.Reque
 		l2Nodes, err := r.findNodesOnSameL2(v1node)
 		if err != nil {
 			log.Error(err, "Could not find nodes on same L2")
-			return ctrl.Result{}, err
+			requeue = 1 * time.Minute
+			return ctrl.Result{RequeueAfter: requeue}, nil
 		}
 
 		// XXX here we should filter for running nodes
 		// i.e. we should have at least 1 running node before suspending it
 		if len(l2Nodes) == 0 {
 			log.Info("No nodes on the same L2 as our node")
-			return ctrl.Result{RequeueAfter: time.Minute}, nil
+			requeue = 1 * time.Minute
+			return ctrl.Result{RequeueAfter: requeue}, nil
 		}
 
 		candidateNode := getCandidateNodeFromList(l2Nodes)
@@ -350,18 +464,32 @@ func (r *AnesthesiaNodeReconciler) Reconcile(ctx context.Context, req ctrl.Reque
 		err = r.ensureWolJob(woljob)
 		if err != nil {
 			log.Error(err, "Could not ensure WOL Job")
-			return ctrl.Result{RequeueAfter: 2 * time.Minute}, err
+			return ctrl.Result{RequeueAfter: 2 * time.Minute}, err // XXX Maybe we need to fail here ?
 		}
 		requeue = 2 * time.Minute
 
 	} else if node.Spec.State == "off" && nodeAlive {
+		// We have to cordon + drain the node, and then shut it down
+		// XXX we need another if here with cordon -> shut down
 		newActualState = "shutting-down"
 		// XXX make sure that we will still have a node alive for WOL
-		// TODO
+
+		// XXX cleanup old shutdown pods
+		err := r.removeOldShutdownPods(&node)
+
+		pod := shutdownPod(&node)
+		err = r.ensurePod(pod)
+		if err != nil {
+			log.Error(err, "Could not create shutdown Pod")
+			return ctrl.Result{RequeueAfter: 2 * time.Minute}, err // XXX maybe we need to fail here
+		}
+		log.Info("Created shutdown pod")
+
+		requeue = 2 * time.Minute
 
 	} else if node.Spec.State == "off" && !nodeAlive {
 		newActualState = "halted"
-		// TODO
+		log.Info("Nothing new to do here !")
 
 	} else {
 		err := fmt.Errorf("Impossible state %v", node.Spec.State)
@@ -369,10 +497,13 @@ func (r *AnesthesiaNodeReconciler) Reconcile(ctx context.Context, req ctrl.Reque
 	}
 
 	if node.Status.ActualState != newActualState {
-		node.Status.ActualState = newActualState
+		log.Info(fmt.Sprintf("Will update state %v->%v", node.Status.ActualState, newActualState))
 
-		log.Info("Will update state")
-		r.Status().Update(ctx, &node)
+		node.Status.ActualState = newActualState
+		err := r.Status().Update(ctx, &node)
+		if err != nil {
+			log.Error(err, "Could not update state")
+		}
 	}
 
 	// TODO(user): your logic here