diff --git a/client/allocrunner/taskrunner/lazy_handle.go b/client/allocrunner/taskrunner/lazy_handle.go
new file mode 100644
index 000000000..5dfd1c39a
--- /dev/null
+++ b/client/allocrunner/taskrunner/lazy_handle.go
@@ -0,0 +1,142 @@
+package taskrunner
+
+import (
+	"fmt"
+	"sync"
+	"time"
+
+	log "github.com/hashicorp/go-hclog"
+	cstructs "github.com/hashicorp/nomad/client/structs"
+	bstructs "github.com/hashicorp/nomad/plugins/base/structs"
+)
+
+const (
+	// retrieveBackoffBaseline is the baseline time for exponential backoff while
+	// retrieving a handle.
+	retrieveBackoffBaseline = 250 * time.Millisecond
+
+	// retrieveBackoffLimit is the limit of the exponential backoff for
+	// retrieving a handle.
+	retrieveBackoffLimit = 5 * time.Second
+
+	// retrieveFailureLimit is how many times we will attempt to retireve a
+	// handle giving up and potentially leaking resources.
+	retrieveFailureLimit = 5
+)
+
+// retrieveHandleFn is used to retrieve the latest driver handle
+type retrieveHandleFn func() *DriverHandle
+
+// LazyHandle is used to front calls to a DriverHandle where it is expected the
+// existing handle may no longer be valid because the backing plugin has
+// shutdown. LazyHandle detects the plugin shutting down and retrieves a new
+// handle so that the consumer does not need to worry whether the handle is to
+// the latest driver instance.
+type LazyHandle struct {
+	// retrieveHandle is used to retrieve the latest handle
+	retrieveHandle retrieveHandleFn
+
+	// h is the current handle and may be nil
+	h *DriverHandle
+
+	logger log.Logger
+	sync.Mutex
+}
+
+// NewLazyHandle takes the function to recieve the latest handle and a logger
+// and returns a LazyHandle
+func NewLazyHandle(fn retrieveHandleFn, logger log.Logger) *LazyHandle {
+	return &LazyHandle{
+		retrieveHandle: fn,
+		h:              fn(),
+		logger:         logger.Named("lazy_handle"),
+	}
+}
+
+// getHandle returns the current handle or retrieves a new one
+func (l *LazyHandle) getHandle() (*DriverHandle, error) {
+	l.Lock()
+	defer l.Unlock()
+
+	if l.h != nil {
+		return l.h, nil
+	}
+
+	return l.refreshHandleLocked()
+}
+
+// refreshHandle retrieves a new handle
+func (l *LazyHandle) refreshHandle() (*DriverHandle, error) {
+	l.Lock()
+	defer l.Unlock()
+	return l.refreshHandleLocked()
+}
+
+// refreshHandleLocked retrieves a new handle and should be called with the lock
+// held. It will retry to give the client time to restart the driver and restore
+// the handle.
+func (l *LazyHandle) refreshHandleLocked() (*DriverHandle, error) {
+	for i := 0; i < retrieveFailureLimit; i++ {
+		l.h = l.retrieveHandle()
+		if l.h != nil {
+			return l.h, nil
+		}
+
+		// Calculate the new backoff
+		backoff := (1 << (2 * uint64(i))) * killBackoffBaseline
+		if backoff > killBackoffLimit {
+			backoff = killBackoffLimit
+		}
+
+		l.logger.Debug("failed to retrieve handle", "backoff", backoff)
+		time.Sleep(backoff)
+	}
+
+	return nil, fmt.Errorf("no driver handle")
+}
+
+func (l *LazyHandle) Exec(timeout time.Duration, cmd string, args []string) ([]byte, int, error) {
+	h, err := l.getHandle()
+	if err != nil {
+		return nil, 0, err
+	}
+
+	// Only retry once
+	first := true
+
+TRY:
+	out, c, err := h.Exec(timeout, cmd, args)
+	if err == bstructs.ErrPluginShutdown && first {
+		first = false
+
+		h, err = l.refreshHandle()
+		if err == nil {
+			goto TRY
+		}
+	}
+
+	return out, c, err
+}
+
+func (l *LazyHandle) Stats() (*cstructs.TaskResourceUsage, error) {
+	h, err := l.getHandle()
+	if err != nil {
+		return nil, err
+	}
+
+	// Only retry once
+	first := true
+
+TRY:
+	out, err := h.Stats()
+	if err == bstructs.ErrPluginShutdown && first {
+		first = false
+
+		h, err = l.refreshHandle()
+		if err == nil {
+			goto TRY
+		}
+	}
+
+	return out, err
+}
diff --git a/client/allocrunner/taskrunner/stats_hook.go b/client/allocrunner/taskrunner/stats_hook.go
index 7343df5bc..31100f471 100644
--- a/client/allocrunner/taskrunner/stats_hook.go
+++ b/client/allocrunner/taskrunner/stats_hook.go
@@ -99,13 +99,15 @@ func (h *statsHook) collectResourceUsageStats(handle interfaces.DriverStats, sto
 					return
 				}
 
-				// We do not log when the plugin is shutdown as this is simply a
-				// race between the stopCollection channel being closed and calling
-				// Stats on the handle.
+				// We do not log when the plugin is shutdown as this is either:
+				// - A race between the stopCollection channel being closed and
+				//	 calling Stats on the handle.
+				// = The driver plugin has unexpectedly exited
+				//
+				// In either case sleeping and trying again or returning based
+				// on the stop channel is the correct behavior
 				if err != bstructs.ErrPluginShutdown {
 					h.logger.Debug("error fetching stats of task", "error", err)
-				} else {
-					// TODO(alex) this breaks if the handle dies
 				}
 
 				continue
diff --git a/client/allocrunner/taskrunner/task_runner_hooks.go b/client/allocrunner/taskrunner/task_runner_hooks.go
index 6aa0da23a..8063630ee 100644
--- a/client/allocrunner/taskrunner/task_runner_hooks.go
+++ b/client/allocrunner/taskrunner/task_runner_hooks.go
@@ -254,6 +254,10 @@ func (tr *TaskRunner) poststart() error {
 	handle := tr.getDriverHandle()
 	net := handle.Network()
 
+	// Pass the lazy handle to the hooks so even if the driver exits and we
+	// launch a new one (external plugin), the handle will refresh.
+	lazyHandle := NewLazyHandle(tr.getDriverHandle, tr.logger)
+
 	var merr multierror.Error
 	for _, hook := range tr.runnerHooks {
 		post, ok := hook.(interfaces.TaskPoststartHook)
@@ -269,9 +273,9 @@ func (tr *TaskRunner) poststart() error {
 		}
 
 		req := interfaces.TaskPoststartRequest{
-			DriverExec:    handle,
+			DriverExec:    lazyHandle,
 			DriverNetwork: net,
-			DriverStats:   handle,
+			DriverStats:   lazyHandle,
 			TaskEnv:       tr.envBuilder.Build(),
 		}
 		var resp interfaces.TaskPoststartResponse