diff --git a/crates/agent_servers/src/acp.rs b/crates/agent_servers/src/acp.rs
index f3bf1ba82b5a00..4789ec012c42dd 100644
--- a/crates/agent_servers/src/acp.rs
+++ b/crates/agent_servers/src/acp.rs
@@ -669,9 +669,12 @@ const MINIMUM_SUPPORTED_VERSION: ProtocolVersion = ProtocolVersion::V1;
 /// dispatch queue via `dispatch_tx`, where they are handled by the
 /// `handle_*` functions on a GPUI context. The returned future drives the
 /// connection and completes when the transport closes; callers are expected
-/// to spawn it on a background executor and hold the task for the lifetime
-/// of the connection. The `connection_tx` oneshot receives the
-/// `ConnectionTo<Agent>` handle as soon as the builder runs its `main_fn`.
+/// to poll it in the background and hold the task for the lifetime of the
+/// connection. In unoptimized builds each inbound dispatch needs ~0.5 MiB
+/// of stack, so poll it on a thread with room to spare (macOS GCD workers'
+/// 512 KiB is not enough — see `AcpConnection::stdio`). The `connection_tx`
+/// oneshot receives the `ConnectionTo<Agent>` handle as soon as the builder
+/// runs its `main_fn`.
 fn connect_client_future(
     name: &'static str,
     transport: impl agent_client_protocol::ConnectTo<Client> + 'static,
@@ -925,17 +928,25 @@ impl AcpConnection {
         });
 
         // `connect_client_future` installs the production handler set and
-        // hands us back both the connection-future (to run on a background
-        // executor) and a oneshot receiver that produces the
-        // `ConnectionTo<Agent>` once the transport handshake is ready.
+        // hands us back both the connection-future and a oneshot receiver
+        // that produces the `ConnectionTo<Agent>` once the transport
+        // handshake is ready. The future must be polled on a dedicated
+        // thread rather than via `background_spawn`: in unoptimized builds
+        // its dispatch chain needs ~0.5 MiB of stack per inbound message,
+        // which overflows the fixed 512 KiB stacks of the GCD workers that
+        // poll background tasks on macOS, crashing dev builds as soon as an
+        // agent sends its first message. See `spawn_dedicated` for the
+        // stack guarantee that makes the dedicated thread sufficient.
         let (connection_tx, connection_rx) = futures::channel::oneshot::channel();
         let connection_future =
             connect_client_future("zed", transport, dispatch_tx.clone(), connection_tx);
-        let io_task = cx.background_spawn(async move {
-            if let Err(err) = connection_future.await {
-                log::error!("ACP connection error: {err}");
-            }
-        });
+        let io_task = cx
+            .background_executor()
+            .spawn_dedicated(move |_executor| async move {
+                if let Err(err) = connection_future.await {
+                    log::error!("ACP connection error: {err}");
+                }
+            });
 
         let connection_rx = async move {
             connection_rx
diff --git a/crates/gpui/src/executor.rs b/crates/gpui/src/executor.rs
index 92afc696d4536e..785a58e95d89ca 100644
--- a/crates/gpui/src/executor.rs
+++ b/crates/gpui/src/executor.rs
@@ -84,6 +84,27 @@ impl BackgroundExecutor {
         self.inner.clone()
     }
 
+    /// Spawn a closure on a fresh session pinned to its own [`SchedulerLocalExecutor`].
+    /// The closure runs on a new OS thread under the platform scheduler, or on
+    /// the test scheduler's loop in tests.
+    ///
+    /// Prefer this over [`Self::spawn`] for futures whose polls need more stack
+    /// than shared background threads guarantee. Dedicated threads get the
+    /// standard library's default 2 MiB, while `spawn` polls futures on
+    /// whatever threads the platform dispatcher provides — on macOS those are
+    /// GCD workers whose stacks are fixed at 512 KiB by the kernel (see `PTH_DEFAULT_STACKSIZE` in
+    /// <https://github.com/apple-oss-distributions/libpthread/blob/42d026df5b07825070f60134b980a1ec2552dfee/kern/kern_internal.h#L154>),
+    /// the tightest background-stack budget of any platform.
+    #[track_caller]
+    pub fn spawn_dedicated<F, Fut>(&self, f: F) -> Task<Fut::Output>
+    where
+        F: FnOnce(SchedulerLocalExecutor) -> Fut + Send + 'static,
+        Fut: Future + 'static,
+        Fut::Output: Send + Sync + 'static,
+    {
+        self.inner.spawn_dedicated(f)
+    }
+
     /// Enqueues the given future to be run to completion on a background thread.
     #[track_caller]
     pub fn spawn<R>(&self, future: impl Future<Output = R> + Send + 'static) -> Task<R>