diff --git a/libcoz/profiler.cpp b/libcoz/profiler.cpp
index 81c53a3..38d9fa5 100644
--- a/libcoz/profiler.cpp
+++ b/libcoz/profiler.cpp
@@ -81,7 +81,7 @@ void profiler::startup(const string& outfile,
 
   // Begin sampling in the main thread
   thread_state* state = add_thread();
-  REQUIRE(state) << "Failed to add thread state";
+  REQUIRE(state) << "Failed to add thread";
   begin_sampling(state);
 }
 
@@ -300,7 +300,11 @@ void profiler::shutdown() {
 }
 
 thread_state* profiler::add_thread() {
-  return _thread_states.insert(gettid());
+  thread_state* inserted = _thread_states.insert(gettid());
+  if (inserted != nullptr) {
+    _num_threads_running += 1;
+  }
+  return inserted;
 }
 
 thread_state* profiler::get_thread_state() {
@@ -309,6 +313,7 @@ thread_state* profiler::get_thread_state() {
 
 void profiler::remove_thread() {
   _thread_states.remove(gettid());
+  _num_threads_running -= 1;
 }
 
 /**
@@ -318,7 +323,7 @@ void* profiler::start_thread(void* p) {
   thread_start_arg* arg = reinterpret_cast<thread_start_arg*>(p);
 
   thread_state* state = get_instance().add_thread();
-  REQUIRE(state) << "Failed to add thread state";
+  REQUIRE(state) << "Failed to add thread";
 
   state->local_delay = arg->_parent_delay_time;
 
diff --git a/libcoz/profiler.h b/libcoz/profiler.h
index 0998a22..d71dbd8 100644
--- a/libcoz/profiler.h
+++ b/libcoz/profiler.h
@@ -132,6 +132,10 @@ class profiler {
   /// Force threads to catch up on delays, and stop sampling before the thread exits
   void handle_pthread_exit(void* result) __attribute__((noreturn)) {
     end_sampling();
+    // If no more threads being sampled, shut down the profiler
+    if (_num_threads_running == 0) {
+      shutdown();
+    }
     real::pthread_exit(result);
     abort(); // Silence g++ warning about noreturn
   }
@@ -223,6 +227,7 @@ class profiler {
   spinlock _latency_points_lock;  //< Spinlock that protects the latency points map
 
   static_map<pid_t, thread_state> _thread_states;   //< Map from thread IDs to thread-local state
+  uint64_t _num_threads_running;                    //< Number of threads that are currently being sampled
 
   std::atomic<bool> _experiment_active; //< Is an experiment running?
   std::atomic<size_t> _global_delay;    //< The global delay time required