diff --git a/agent/sandbox/Makefile b/agent/sandbox/Makefile
index bcd5eef76a9..f7aa3ac8117 100644
--- a/agent/sandbox/Makefile
+++ b/agent/sandbox/Makefile
@@ -102,14 +102,23 @@ logs:
 clean:
 	@echo "🧹 Cleaning all containers and volumes..."
 	@docker compose down -v || true
-	@if [ -f .env ]; then \
-		source .env && \
-		for i in $$(seq 0 $$((SANDBOX_EXECUTOR_MANAGER_POOL_SIZE - 1))); do \
-			echo "🧹 Deleting sandbox_python_$$i..." && \
-			docker rm -f sandbox_python_$$i 2>/dev/null || true && \
-			echo "🧹 Deleting sandbox_nodejs_$$i..." && \
-			docker rm -f sandbox_nodejs_$$i 2>/dev/null || true; \
-		done; \
+	@# Same discovery as scripts/stop.sh: the containers that exist, not a range derived from
+	@# SANDBOX_EXECUTOR_MANAGER_POOL_SIZE, which misses a pool created at a different size, skips
+	@# cleanup entirely when .env is gone, and never covers a container that failed to join the
+	@# executor-manager queue.
+	@# A failed `docker ps` and an empty result both collapse to an empty string, and only the
+	@# first means the containers are still running. Tolerate grep's no-match exit, not docker's.
+	@if ! all_containers="$$(docker ps -a --format '{{.Names}}')"; then \
+		echo "❌ Could not list Docker containers; sandbox containers were NOT removed" >&2; \
+		exit 1; \
+	fi; \
+	sandbox_containers="$$(printf '%s\n' "$$all_containers" | \
+		grep -E '^sandbox_(python|nodejs)_[0-9]+$$' || true)"; \
+	if [ -n "$$sandbox_containers" ]; then \
+		while IFS= read -r container; do \
+			echo "🧹 Deleting $$container..."; \
+			docker rm -f "$$container" >/dev/null 2>&1 || true; \
+		done <<<"$$sandbox_containers"; \
 	else \
-		echo "⚠️ .env not found, skipping container cleanup"; \
+		echo "✅ No sandbox containers found"; \
 	fi
diff --git a/agent/sandbox/scripts/stop.sh b/agent/sandbox/scripts/stop.sh
index 51bd2b6e93b..8e2ec02ff23 100755
--- a/agent/sandbox/scripts/stop.sh
+++ b/agent/sandbox/scripts/stop.sh
@@ -24,17 +24,28 @@ echo "🛑 Stopping all services..."
 docker compose down
 
 echo "🧹 Deleting sandbox containers..."
-if [ -f .env ]; then
-  source .env
-  for i in $(seq 0 $((SANDBOX_EXECUTOR_MANAGER_POOL_SIZE - 1))); do
-    echo "🧹 Deleting sandbox_python_$i..."
-    docker rm -f "sandbox_python_$i" >/dev/null 2>&1 || true
-
-    echo "🧹 Deleting sandbox_nodejs_$i..."
-    docker rm -f "sandbox_nodejs_$i" >/dev/null 2>&1 || true
-  done
+# Remove the sandbox containers that actually exist instead of walking a range derived from
+# SANDBOX_EXECUTOR_MANAGER_POOL_SIZE. A pool created at one size and torn down after .env changed
+# left every container above the new size running; a missing .env skipped cleanup altogether; and a
+# container that never joined the executor-manager queue was never in the range to begin with.
+# The pattern is anchored to the generated pool names, so an unrelated container that merely
+# contains "sandbox_python_" in its name is not touched.
+# `docker ps` failing and finding nothing look identical once both become an empty string, and
+# the difference matters: the first leaves every sandbox container running while reporting
+# success. Only grep's no-match exit is tolerated here.
+if ! all_containers="$(docker ps -a --format '{{.Names}}')"; then
+  echo "❌ Could not list Docker containers; sandbox containers were NOT removed" >&2
+  exit 1
+fi
+sandbox_containers="$(printf '%s\n' "$all_containers" |
+  grep -E '^sandbox_(python|nodejs)_[0-9]+$' || true)"
+if [ -n "$sandbox_containers" ]; then
+  while IFS= read -r container; do
+    echo "🧹 Deleting $container..."
+    docker rm -f "$container" >/dev/null 2>&1 || true
+  done <<<"$sandbox_containers"
 else
-  echo "⚠️ .env not found, skipping container cleanup"
+  echo "✅ No sandbox containers found"
 fi
 
 echo "✅ Stopping and cleanup complete"