diff --git a/backend/apps/agents/core/error_classify.py b/backend/apps/agents/core/error_classify.py
index cc3008b9..27e915d3 100644
--- a/backend/apps/agents/core/error_classify.py
+++ b/backend/apps/agents/core/error_classify.py
@@ -1,7 +1,6 @@
 import re
 from typing import Optional, Tuple
 
-import anthropic
 import httpx
 from typeguard import typechecked
 
@@ -207,9 +206,19 @@ def parse_retry_after(exc: BaseException, extra_text: str = "") -> int | None:
 # anthropic.APIConnectionError stringifies to the bare "Connection error.", so the patterns above
 # score it NON-transient and one network hiccup throws away a whole run (measured live, twice). A
 # transport failure is transient by construction, so classify by TYPE, which no rewording breaks.
-P_TRANSIENT_EXC_TYPES: Tuple[type, ...] = (
-    anthropic.APIConnectionError, anthropic.InternalServerError,  # APITimeoutError subclasses the first
-    httpx.TransportError, ConnectionError, TimeoutError)          # connect/read/pool timeouts, protocol errors
+# Built lazily: importing the anthropic SDK at module scope cost 224ms of every backend boot.
+p_transient_exc_types: Optional[Tuple[type, ...]] = None
+
+
+def p_get_transient_exc_types() -> Tuple[type, ...]:
+    global p_transient_exc_types
+    if p_transient_exc_types is None:
+        import anthropic
+
+        p_transient_exc_types = (
+            anthropic.APIConnectionError, anthropic.InternalServerError,  # APITimeoutError subclasses the first
+            httpx.TransportError, ConnectionError, TimeoutError)          # connect/read/pool timeouts, protocol errors
+    return p_transient_exc_types
 
 
 @typechecked
@@ -222,7 +231,7 @@ def is_transient_capacity_error(exc: BaseException, extra_text: str = "") -> boo
     if is_context_overflow_error(exc, extra_text):
         return False
     # Ahead of the empty-string bail on purpose: what the exception IS doesn't depend on whether it bothered to say anything.
-    if isinstance(exc, P_TRANSIENT_EXC_TYPES):
+    if isinstance(exc, p_get_transient_exc_types()):
         return True
     if not combined:
         return False
diff --git a/backend/apps/agents/manager/permissions/gate_hooks.py b/backend/apps/agents/manager/permissions/gate_hooks.py
index d85b683b..c83725fb 100644
--- a/backend/apps/agents/manager/permissions/gate_hooks.py
+++ b/backend/apps/agents/manager/permissions/gate_hooks.py
@@ -11,7 +11,13 @@ import time
 from typing import Dict, Optional, Union
 
 from typeguard import typechecked
-from claude_agent_sdk import PermissionResultAllow, PermissionResultDeny
+# Runtime aliases stay `object` so the 350ms claude_agent_sdk+mcp chain stays off the boot graph; the hook body re-imports the real classes at call time, when the SDK is already resident.
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from claude_agent_sdk import PermissionResultAllow, PermissionResultDeny
+else:
+    PermissionResultAllow = PermissionResultDeny = object
 
 from backend.apps.agents.core.ws_manager import ws_manager
 from backend.apps.settings.settings import load_settings
@@ -36,6 +42,8 @@ logger = logging.getLogger(__name__)
 async def can_use_tool(
     ctx: HookContext, tool_name: str, input_data: object, context: object
 ) -> Union[PermissionResultAllow, PermissionResultDeny]:
+    from claude_agent_sdk import PermissionResultAllow, PermissionResultDeny
+
     if is_claude_schedule_skill(tool_name, input_data):
         note_tool_used(ctx.session_id, tool_name, False)
         return PermissionResultDeny(
diff --git a/backend/apps/agents/manager/streaming/handle_assistant_message.py b/backend/apps/agents/manager/streaming/handle_assistant_message.py
index e7a4dfc4..fb9d0c7b 100644
--- a/backend/apps/agents/manager/streaming/handle_assistant_message.py
+++ b/backend/apps/agents/manager/streaming/handle_assistant_message.py
@@ -18,11 +18,13 @@ from backend.apps.agents.manager.streaming.upsert_message import upsert_message
 from backend.apps.agents.manager.streaming.PartialReply import PartialReply
 from backend.apps.agents.manager.streaming import thinking as thinking_mod
 
-try:
+# The block types drive isinstance DISPATCH, so they must be real at runtime; imported inside the handler because by stream time the SDK is already resident (the turn's presence check imported it), keeping the 350ms sdk+mcp chain off the boot graph.
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
     from claude_agent_sdk import AssistantMessage
-    from claude_agent_sdk.types import ThinkingBlock, TextBlock, ToolUseBlock
-except ImportError:  # the SDK is optional at runtime (mock mode); keep this module importable
-    AssistantMessage = ThinkingBlock = TextBlock = ToolUseBlock = object  # type: ignore
+else:
+    AssistantMessage = object
 
 
 @typechecked
@@ -35,6 +37,8 @@ async def handle_assistant_message(
     live_partial: Dict[str, PartialReply],
     sessions: Dict[str, AgentSession],
 ) -> None:
+    from claude_agent_sdk.types import ThinkingBlock, TextBlock, ToolUseBlock
+
     content_parts = []
     new_thinking_parts = []
     tool_uses = []
diff --git a/backend/apps/agents/manager/streaming/handle_result_message.py b/backend/apps/agents/manager/streaming/handle_result_message.py
index fae05a87..7fc3a4cb 100644
--- a/backend/apps/agents/manager/streaming/handle_result_message.py
+++ b/backend/apps/agents/manager/streaming/handle_result_message.py
@@ -15,10 +15,13 @@ from backend.apps.agents.core.ws_manager import ws_manager
 from backend.apps.agents.manager.streaming.state import ThinkingState, TurnState
 from backend.apps.agents.manager.streaming import thinking as thinking_mod
 
-try:
+# Annotation-only here (no isinstance dispatch), so the runtime symbol can stay `object` and the SDK chain stays off the boot import graph.
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
     from claude_agent_sdk import ResultMessage
-except ImportError:  # the SDK is optional at runtime (mock mode); keep this module importable
-    ResultMessage = object  # type: ignore
+else:
+    ResultMessage = object
 
 logger = logging.getLogger(__name__)
 
diff --git a/backend/apps/agents/manager/streaming/handle_stream_event.py b/backend/apps/agents/manager/streaming/handle_stream_event.py
index 17d5a626..33d91157 100644
--- a/backend/apps/agents/manager/streaming/handle_stream_event.py
+++ b/backend/apps/agents/manager/streaming/handle_stream_event.py
@@ -15,10 +15,14 @@ from backend.apps.agents.core.ws_manager import ws_manager
 from backend.apps.agents.manager.streaming.state import ThinkingState, TurnState
 from backend.apps.agents.manager.streaming.PartialReply import PartialReply
 
-try:
+# Runtime annotation stays `object` (the old ImportError fallback already admitted that); the real
+# type lives behind TYPE_CHECKING so importing this module stops paying the 350ms claude_agent_sdk+mcp chain at boot.
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
     from claude_agent_sdk.types import StreamEvent
-except ImportError:  # the SDK is optional at runtime (mock mode); keep this module importable
-    StreamEvent = object  # type: ignore
+else:
+    StreamEvent = object
 
 
 @typechecked