diff --git a/Cargo.toml b/Cargo.toml
index 86bf1906..453f0f2b 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -42,6 +42,7 @@ quad-gl = { git = "https://github.com/not-fl3/quad-gl", branch = "v0.4" }
 
 [dev-dependencies]
 dolly = "*"
+tokio = { version = "1.41.1", features = ["fs", "rt"] }
 
 #[patch."https://github.com/not-fl3/quad-gl"]
 #quad-gl = { path = '../quad-gl' }
diff --git a/examples/other_async_executors.rs b/examples/other_async_executors.rs
new file mode 100644
index 00000000..c54d83ad
--- /dev/null
+++ b/examples/other_async_executors.rs
@@ -0,0 +1,23 @@
+use macroquad::color::*;
+
+async fn game(ctx: macroquad::Context) {
+    let rt = tokio::runtime::Builder::new_current_thread()
+        .build()
+        .unwrap();
+    let _guard = rt.enter();
+    for _ in 0..3 {
+        ctx.clear_screen(WHITE);
+
+        eprintln!("now for some tokio business");
+
+        let file = tokio::fs::File::open("examples/ferris.png").await.unwrap();
+
+        eprintln!("tokio file loaded");
+
+        ctx.next_frame().await;
+    }
+}
+
+fn main() {
+    macroquad::start(Default::default(), |ctx| game(ctx));
+}
diff --git a/src/compat.rs b/src/compat.rs
index a41ff291..847eb4f6 100644
--- a/src/compat.rs
+++ b/src/compat.rs
@@ -1,15 +1,18 @@
 //! miniquad-0.4 emulation
 
-pub use crate::window::next_frame;
 pub use quad_gl::color::*;
 
 use std::{
     cell::RefCell,
     sync::{Arc, Mutex},
+    task::Waker,
 };
 
+use crate::exec::FrameFuture;
+
 pub struct CompatContext {
     quad_ctx: Arc<Mutex<Box<miniquad::Context>>>,
+    frame_wakers: Arc<Mutex<Vec<Waker>>>,
     canvas: quad_gl::sprite_batcher::SpriteBatcher,
 }
 
@@ -17,15 +20,26 @@ thread_local! {
     pub static CTX: RefCell<Option<CompatContext>> = { RefCell::new(None) };
 }
 
-fn with_ctx<F: Fn(&mut CompatContext)>(f: F) {
-    CTX.with_borrow_mut(|v| f(v.as_mut().unwrap()));
+pub fn next_frame() -> FrameFuture {
+    with_ctx(|ctx| FrameFuture {
+        frame_wakers: Some(ctx.frame_wakers.clone()),
+    })
+}
+
+fn with_ctx<T, F: Fn(&mut CompatContext) -> T>(f: F) -> T {
+    CTX.with_borrow_mut(|v| f(v.as_mut().unwrap()))
 }
 pub fn init_compat_mode(ctx: &crate::Context) {
     let canvas = ctx.new_canvas();
     let quad_ctx = ctx.quad_ctx.clone();
+    let frame_wakers = ctx.frame_wakers.clone();
 
     CTX.with_borrow_mut(|v| {
-        *v = Some(CompatContext { quad_ctx, canvas });
+        *v = Some(CompatContext {
+            quad_ctx,
+            canvas,
+            frame_wakers,
+        });
     });
 }
 
diff --git a/src/exec.rs b/src/exec.rs
index aed90fc9..931a1989 100644
--- a/src/exec.rs
+++ b/src/exec.rs
@@ -1,28 +1,31 @@
+use std::collections::{HashMap, VecDeque};
 use std::future::Future;
 use std::pin::Pin;
+use std::sync::atomic::AtomicU64;
 use std::sync::{Arc, Mutex};
-use std::task::{Context, Poll, RawWaker, RawWakerVTable, Waker};
+use std::task::{Context, Poll, RawWaker, RawWakerVTable, Wake, Waker};
 
 use crate::Error;
 
-// Returns Pending as long as its inner bool is false.
-#[derive(Default)]
+// Returns Pending once and finishes immediately once woken.
 pub struct FrameFuture {
-    done: bool,
+    pub frame_wakers: Option<Arc<Mutex<Vec<Waker>>>>,
 }
 
 impl Future for FrameFuture {
     type Output = ();
 
-    fn poll(mut self: Pin<&mut Self>, _context: &mut Context) -> Poll<Self::Output> {
-        if self.done {
+    fn poll(mut self: Pin<&mut Self>, context: &mut Context) -> Poll<Self::Output> {
+        if let Some(wakers) = self.frame_wakers.take() {
+            wakers.lock().unwrap().push(context.waker().clone());
+            eprintln!("frame future pend");
+            Poll::Pending
+        } else {
             // We were told to step, meaning this future gets destroyed and we run
             // the main future until we call next_frame again and end up in this poll
             // function again.
+            eprintln!("frame future ready");
             Poll::Ready(())
-        } else {
-            self.done = true;
-            Poll::Pending
         }
     }
 }
@@ -44,33 +47,115 @@ impl Future for FileLoadingFuture {
     }
 }
 
-fn waker() -> Waker {
+fn waker(inner: Arc<Inner<'_>>, id: u64) -> Waker {
+    // Cannot use the `Wake` trait, because `Inner` has a lifetime that
+    // cannot be used to generate a raw waker.
     unsafe fn clone(data: *const ()) -> RawWaker {
+        Arc::increment_strong_count(data.cast::<InnerWaker<'_>>());
         RawWaker::new(data, &VTABLE)
     }
-    unsafe fn wake(_data: *const ()) {
-        panic!(
-            "macroquad does not support waking futures, please use coroutines, \
-            otherwise your pending future will block until the next frame"
-        )
+    unsafe fn wake(data: *const ()) {
+        // Even though we generate an unconstrained lifetime here, that's no issue,
+        // as we just move data that has a lifetime from one collection to another.
+        let data = Arc::<InnerWaker<'_>>::from_raw(data.cast());
+        data.wake();
     }
     unsafe fn wake_by_ref(data: *const ()) {
-        wake(data)
+        let data = &*data.cast::<InnerWaker<'_>>();
+        data.wake();
     }
-    unsafe fn drop(_data: *const ()) {
-        // Nothing to do
+    unsafe fn drop(data: *const ()) {
+        Arc::<InnerWaker<'_>>::from_raw(data.cast());
     }
     const VTABLE: RawWakerVTable = RawWakerVTable::new(clone, wake, wake_by_ref, drop);
-    let raw_waker = RawWaker::new(std::ptr::null(), &VTABLE);
+    let raw_waker = RawWaker::new(
+        Arc::into_raw(Arc::new(InnerWaker { inner, id })).cast(),
+        &VTABLE,
+    );
     unsafe { Waker::from_raw(raw_waker) }
 }
 
-/// returns Some(T) if future is done, None if it would block
-pub(crate) fn resume<T>(future: &mut Pin<Box<dyn Future<Output = T>>>) -> Option<T> {
-    let waker = waker();
-    let mut futures_context = std::task::Context::from_waker(&waker);
-    match future.as_mut().poll(&mut futures_context) {
-        Poll::Ready(v) => Some(v),
-        Poll::Pending => None,
+type LocalFuture<'a> = Pin<Box<dyn Future<Output = ()> + 'a>>;
+
+/// A simple unoptimized, single threaded executor.
+pub struct Executor<'a> {
+    inner: Arc<Inner<'a>>,
+    pub frame_wakers: Arc<Mutex<Vec<Waker>>>,
+}
+
+#[derive(Default)]
+struct Inner<'a> {
+    // FIXME: use a thread safe queue that doesn't need to lock the read end to
+    // push to the write end.
+    active_futures: Mutex<VecDeque<LocalFuture<'a>>>,
+    waiting_futures: Mutex<HashMap<u64, LocalFuture<'a>>>,
+    next_wait_id: AtomicU64,
+}
+
+struct InnerWaker<'a> {
+    id: u64,
+    inner: Arc<Inner<'a>>,
+}
+
+impl InnerWaker<'_> {
+    fn wake(&self) {
+        let future = self
+            .inner
+            .waiting_futures
+            .lock()
+            .unwrap()
+            .remove(&self.id)
+            .expect("already woken");
+        self.inner.active_futures.lock().unwrap().push_back(future);
+    }
+}
+
+impl<'a> Executor<'a> {
+    pub fn new(frame_wakers: Arc<Mutex<Vec<Waker>>>) -> Self {
+        Self {
+            frame_wakers,
+            inner: Default::default(),
+        }
+    }
+
+    /// Returns `true` if all futures have finished.
+    pub fn is_empty(&self) -> bool {
+        self.inner.active_futures.lock().unwrap().is_empty()
+            && self.inner.waiting_futures.lock().unwrap().is_empty()
+    }
+
+    /// Runs one future until it returns `Pending`.
+    /// Returns `true` if anything was run.
+    pub fn tick(&self) -> bool {
+        let Some(mut future) = self.inner.active_futures.lock().unwrap().pop_front() else {
+            return false;
+        };
+        let id = self
+            .inner
+            .next_wait_id
+            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+        let waker = waker(self.inner.clone(), id);
+        let mut futures_context = std::task::Context::from_waker(&waker);
+        match future.as_mut().poll(&mut futures_context) {
+            Poll::Ready(()) => {}
+            Poll::Pending => {
+                let prev = self
+                    .inner
+                    .waiting_futures
+                    .lock()
+                    .unwrap()
+                    .insert(id, future);
+                assert!(prev.is_none());
+            }
+        }
+        true
+    }
+
+    pub fn spawn<Fut: Future<Output = ()> + 'a>(&self, f: Fut) {
+        self.inner
+            .active_futures
+            .lock()
+            .unwrap()
+            .push_back(Box::pin(f));
     }
 }
diff --git a/src/gizmos.rs b/src/gizmos.rs
index a3dea009..c77d0906 100644
--- a/src/gizmos.rs
+++ b/src/gizmos.rs
@@ -1,7 +1,4 @@
-pub use crate::{
-    math::{vec2, vec3, Vec3},
-    window::next_frame,
-};
+pub use crate::math::{vec2, vec3, Vec3};
 pub use quad_gl::{
     color::*,
     draw_calls_batcher::{DrawCallsBatcher, DrawMode, Vertex},
diff --git a/src/lib.rs b/src/lib.rs
index 80c04fee..7fbb8424 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -36,11 +36,13 @@
 //! }
 //!```
 #![allow(warnings)]
+use exec::{Executor, FrameFuture};
 use miniquad::*;
 
 use std::collections::{HashMap, HashSet};
 use std::future::Future;
 use std::pin::Pin;
+use std::task::Waker;
 
 mod exec;
 mod time;
@@ -156,7 +158,7 @@ impl MiniquadInputEvent {
 }
 
 struct Stage {
-    main_future: Option<Pin<Box<dyn Future<Output = ()>>>>,
+    ex: Executor<'static>,
     quad_ctx: Arc<Mutex<Box<miniquad::Context>>>,
     quad_gl: Arc<Mutex<quad_gl::QuadGl>>,
     ui: Arc<Mutex<quad_gl::ui::Ui>>,
@@ -321,18 +323,25 @@ impl EventHandler for Stage {
     fn draw(&mut self) {
         self.time.lock().unwrap().update();
         //let result = maybe_unwind(get_context().unwind, || {
-        if let Some(future) = self.main_future.as_mut() {
-            let _z = telemetry::ZoneGuard::new("user code");
+        let _z = telemetry::ZoneGuard::new("user code");
 
-            if exec::resume(future).is_some() {
-                self.main_future = None;
-                miniquad::window::quit();
-                return;
-            }
+        eprintln!("frame start");
 
-            self.input.lock().unwrap().end_frame();
-            compat::end_frame();
+        for waker in self.ex.frame_wakers.lock().unwrap().drain(..) {
+            waker.wake();
         }
+
+        while self.ex.tick() {}
+
+        eprintln!("frame end");
+
+        if self.ex.is_empty() {
+            miniquad::window::quit();
+            return;
+        }
+
+        self.input.lock().unwrap().end_frame();
+        compat::end_frame();
     }
 
     fn window_restored_event(&mut self) {
@@ -361,6 +370,7 @@ pub struct Context {
     pub input: Arc<Mutex<input::InputContext>>,
     time: Arc<Mutex<time::Time>>,
     ui: Arc<Mutex<quad_gl::ui::Ui>>,
+    frame_wakers: Arc<Mutex<Vec<Waker>>>,
 }
 
 impl Context {
@@ -379,6 +389,7 @@ impl Context {
             input: Arc::new(Mutex::new(input::InputContext::new())),
             ui: Arc::new(Mutex::new(ui)),
             time: Arc::new(Mutex::new(time)),
+            frame_wakers: Default::default(),
         }
     }
 
@@ -462,6 +473,12 @@ impl Context {
     pub fn root_ui<'a>(&'a self) -> impl std::ops::DerefMut<Target = quad_gl::ui::Ui> + 'a {
         self.ui.lock().unwrap()
     }
+
+    pub fn next_frame(&self) -> FrameFuture {
+        FrameFuture {
+            frame_wakers: Some(self.frame_wakers.clone()),
+        }
+    }
 }
 
 pub fn start<F: Fn(Context) -> Fut + 'static, Fut: Future<Output = ()> + 'static>(
@@ -479,7 +496,13 @@ pub fn start<F: Fn(Context) -> Fut + 'static, Fut: Future<Output = ()> + 'static
             quad_gl: ctx.quad_gl.clone(),
             ui: ctx.ui.clone(),
             time: ctx.time.clone(),
-            main_future: Some(Box::pin(future(ctx))),
+            ex: {
+                let ex = Executor::new(ctx.frame_wakers.clone());
+                // TODO: should we wait for coroutines to finish even if the main task finished?
+                // Right now we do, but this is not how macroquad worked previously.
+                ex.spawn(future(ctx));
+                ex
+            },
         })
     });
 }
diff --git a/src/window.rs b/src/window.rs
index 9c49836e..5b681439 100644
--- a/src/window.rs
+++ b/src/window.rs
@@ -7,11 +7,6 @@ pub use miniquad;
 
 pub use miniquad::conf::Conf;
 
-/// Block execution until the next frame.
-pub fn next_frame() -> crate::exec::FrameFuture {
-    crate::exec::FrameFuture::default()
-}
-
 #[doc(hidden)]
 pub fn gl_set_drawcall_buffer_capacity(_max_vertices: usize, _max_indices: usize) {
     // let context = get_context();