diff --git a/backend/apps/agents/manager/run/client_pool.py b/backend/apps/agents/manager/run/client_pool.py
index 02796dbe..17b9f611 100644
--- a/backend/apps/agents/manager/run/client_pool.py
+++ b/backend/apps/agents/manager/run/client_pool.py
@@ -28,7 +28,8 @@ P_NON_BOOT_KEYS = frozenset({"can_use_tool", "stderr", "hooks", "resume", "fork_
 
 
 def persistent_client_enabled() -> bool:
-    return os.environ.get("OSW_TTFT_PERSISTENT_CLIENT") == "1"
+    """Default ON (soak-proven: warm turns 535ms -> 6ms). Kill switch: OPENSWARM_PERSISTENT_CLIENT=0."""
+    return os.environ.get("OPENSWARM_PERSISTENT_CLIENT", "1") != "0"
 
 
 # Per-session field-level digests from the last fingerprint call; lets a mismatch log WHICH boot field drifted (probe-gated diagnostics only).
diff --git a/backend/tests/conftest.py b/backend/tests/conftest.py
index 42ef7c95..c5952c28 100644
--- a/backend/tests/conftest.py
+++ b/backend/tests/conftest.py
@@ -20,8 +20,8 @@ from types import SimpleNamespace
 
 import pytest
 
-# Tests mock claude_agent_sdk.query, not ClaudeSDKClient; an inherited persistent-client flag would route mocked turns onto a REAL CLI spawn and wedge the suite. Pin it off; the persistent path has its own tests (test_client_pool.py + live gates).
-os.environ.pop("OSW_TTFT_PERSISTENT_CLIENT", None)
+# Tests mock claude_agent_sdk.query, not ClaudeSDKClient; the now-default-ON persistent client would route mocked turns onto a REAL CLI spawn and wedge the suite. Pin it OFF explicitly; the persistent path has its own tests (test_client_pool.py + live gates).
+os.environ["OPENSWARM_PERSISTENT_CLIENT"] = "0"
 
 
 @pytest.fixture(autouse=True)