diff --git a/src/lib/onboard/sandbox-gpu-create-flow.test.ts b/src/lib/onboard/sandbox-gpu-create-flow.test.ts
index 217880717a1..690005f3810 100644
--- a/src/lib/onboard/sandbox-gpu-create-flow.test.ts
+++ b/src/lib/onboard/sandbox-gpu-create-flow.test.ts
@@ -507,8 +507,10 @@ describe("runSandboxGpuCreateFlow proof authorization", () => {
 });
 
 describe("runSandboxGpuCreateFlow native failure and readiness", () => {
-  it("threads restart-safe startup resource limits on the no-GPU Docker route", async () => {
+  it("defers restart-safe no-GPU recreation until the create process exits (#8720)", async () => {
     const input = createInput();
+    const patch = createPatch();
+    mocks.createDockerGpuSandboxCreatePatch.mockReturnValueOnce(patch);
     input.sandboxGpuConfig = {
       ...input.sandboxGpuConfig,
       mode: "0",
@@ -534,6 +536,17 @@ describe("runSandboxGpuCreateFlow native failure and readiness", () => {
         requiredUlimits: input.requiredUlimits,
       }),
     );
+    expect(mocks.streamSandboxCreate).toHaveBeenCalledWith(
+      "openshell",
+      ["sandbox", "create"],
+      input.sandboxEnv,
+      expect.objectContaining({ waitForReadyTermination: true }),
+    );
+    const onPoll = mocks.streamSandboxCreate.mock.calls[0]?.[3]?.onPoll;
+    expect(onPoll).toBeTypeOf("function");
+    onPoll();
+    expect(patch.maybeApplyDuringCreate).not.toHaveBeenCalled();
+    expect(patch.ensureApplied).toHaveBeenCalledOnce();
   });
 
   it("does not replace a native GPU container solely to persist its startup command", async () => {
diff --git a/src/lib/onboard/sandbox-gpu-create-run-attempt.ts b/src/lib/onboard/sandbox-gpu-create-run-attempt.ts
index 8cecbc64b2a..5b91caa7368 100644
--- a/src/lib/onboard/sandbox-gpu-create-run-attempt.ts
+++ b/src/lib/onboard/sandbox-gpu-create-run-attempt.ts
@@ -128,6 +128,10 @@ export function createSandboxGpuCreateAttemptRunner(
           },
         })
       : null;
+    const persistRestartSafeStartup =
+      input.persistStartupCommand === true && (route !== "native" || hasRequiredUlimits);
+    const deferRestartSafeCutover =
+      !managedLifecycle && !compatibility && persistRestartSafeStartup;
     const runtimePatch =
       managedLifecycle?.patch ??
       createDockerGpuSandboxCreatePatch({
@@ -135,8 +139,7 @@ export function createSandboxGpuCreateAttemptRunner(
         // The startup clone preserves native CDI devices, so DCode can apply its
         // exact required limits without replacing the native GPU envelope.
         // Other native routes are not swapped solely to persist a command.
-        persistStartupCommand:
-          input.persistStartupCommand === true && (route !== "native" || hasRequiredUlimits),
+        persistStartupCommand: persistRestartSafeStartup,
         externalRecreation: false,
         sandboxName: input.sandboxName,
         gpuDevice: input.sandboxGpuConfig.sandboxGpuDevice,
@@ -161,13 +164,16 @@ export function createSandboxGpuCreateAttemptRunner(
           const list = deps.runCaptureOpenshell(["sandbox", "list"], { ignoreError: true });
           return isSandboxReady(list, input.sandboxName);
         },
-        onPoll: () => runtimePatch.maybeApplyDuringCreate(),
+        onPoll: () => {
+          if (!deferRestartSafeCutover) runtimePatch.maybeApplyDuringCreate();
+        },
         readyCheckOutputPatterns: getReadyCheckOutputPatternsForAgent(
           input.terminalAgent,
           input.sandboxEnv,
         ),
         failureCheck: runtimePatch.createFailureMessage,
         traceEvent: addTraceEvent,
+        waitForReadyTermination: deferRestartSafeCutover,
         initialPhase:
           compatibility && (input.prebuild.imageRef || state.compatibilityArgv)
             ? "create"