diff --git a/CHANGELOG.md b/CHANGELOG.md
index b448227a..21800a08 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -221,6 +221,13 @@ Sessions survive and nobody signs in again.
   laptop `http://localhost` counts as one, so this never showed up in development; on a real
   address it does not, and the surface did nothing at all when you pressed send. No message, no
   error. Ids now come from an API with no such restriction.
+- **The first browser action a Bot was ever asked for failed.** Creating a computer and starting it
+  are two calls to Docker, and a name the daemon has not published yet answers the second with a 404.
+  The supervisor treats that as a lost race and rebuilds, which is right, but it went straight back
+  round: the retry landed a millisecond later, saw the same unpublished name, and spent the only
+  other attempt on it. The whole request then failed as Docker being unreachable, the person was told
+  the computer could not be started, and the next message worked. It waits one poll interval before
+  rebuilding now, which is what the health wait already uses for the same question.
 - **A framework Bot asked for a browser action and nothing happened.** `agent-langgraph` ends a run
   when the model calls a tool the surface owns, which is how a tool that lives in the browser is
   supposed to work: the run finishes, the surface acts, and the next run carries the result. But the
diff --git a/supervisor/src/docker.ts b/supervisor/src/docker.ts
index 7b3d551f..898513ed 100644
--- a/supervisor/src/docker.ts
+++ b/supervisor/src/docker.ts
@@ -45,6 +45,13 @@ function statusOf(error: unknown): number | undefined {
   return (error as { statusCode?: number }).statusCode;
 }
 
+/** One poll interval, used both by the health wait and by the retry that follows a lost race. */
+function pause(ms: number): Promise<void> {
+  const { promise, resolve } = Promise.withResolvers<void>();
+  setTimeout(resolve, ms);
+  return promise;
+}
+
 export type ComputerState = {
   botId: string;
   container: string;
@@ -180,7 +187,7 @@ async function waitUntilAnswering(
     } catch {
       // Mid-creation, or gone. The deadline is what ends this.
     }
-    await new Promise((resolve) => setTimeout(resolve, 250));
+    await pause(250);
   }
 }
 
@@ -311,10 +318,22 @@ export async function ensure(
         await docker.getContainer(names.container).start();
       } catch (error) {
         const status = statusOf(error);
-        // Gone between creating it and starting it: a concurrent reset took the container away, or
-        // the create this request lost the race to was itself rolled back. Nothing about the Bot has
-        // changed, so the answer is to build it again rather than to report Docker as unreachable.
-        if (status === 404 && attempt > 1) continue;
+        /*
+         * Gone between creating it and starting it: a concurrent reset took the container away, the
+         * create this request lost the race to was itself rolled back, or the daemon has not yet
+         * published the name this request just created. Nothing about the Bot has changed, so the
+         * answer is to build it again rather than to report Docker as unreachable.
+         *
+         * Paused first, because the retry is the whole budget. Going straight back round arrives
+         * within a millisecond, sees the same not-yet-published name, and spends the second attempt
+         * on the state that failed the first: the first browser action a Bot is ever asked for fails,
+         * and the second one, seconds later, works. One poll interval is what the health wait uses
+         * for the same question.
+         */
+        if (status === 404 && attempt > 1) {
+          await pause(250);
+          continue;
+        }
         // 304 is "already running", which is success for an idempotent verb.
         if (status !== 304) {
           throw new DockerUnavailableError(String(error));