diff --git a/src/band/adapters/crewai.py b/src/band/adapters/crewai.py
index f49924fe1..ba7047123 100644
--- a/src/band/adapters/crewai.py
+++ b/src/band/adapters/crewai.py
@@ -405,7 +405,9 @@ async def _process_message(
             # "already handled" / "act on this now" markers, matches what
             # CrewAI actually does with the input either way.
             prompt = "\n\n".join(sections)
-            result = await self._crewai_agent.kickoff_async(prompt)
+            result = await self._kickoff_with_empty_response_retry(
+                self._crewai_agent, prompt, reply_tracker, room_id
+            )
 
         except Exception as e:
             # An empty response is benign only once some tool ran this turn --
@@ -461,6 +463,48 @@ async def on_cleanup(self, room_id: str) -> None:
             del self._message_history[room_id]
             logger.debug("Room %s: Cleaned up CrewAI session", room_id)
 
+    async def _kickoff_with_empty_response_retry(
+        self,
+        agent: "CrewAIAgent",
+        prompt: str,
+        reply_tracker: ReplyTracker,
+        room_id: str,
+    ) -> Any:
+        """Retry a first-call empty completion once before giving up.
+
+        CrewAI raises the identical ValueError whether the model correctly has
+        nothing left to say (fine once a tool has run -- handled by the
+        caller) or the turn's very first call came back empty. The second case
+        has run no tool yet, so there is nothing to duplicate: one immediate
+        retry absorbs a single-call fluke instead of failing the whole
+        delivery and leaving CrewAI to improvise on a cold redelivery.
+        """
+        try:
+            return await agent.kickoff_async(prompt)
+        except Exception as e:
+            if not (_is_empty_llm_response(e) and not reply_tracker.any_tool_ran):
+                raise
+            logger.info(
+                "Room %s: CrewAI's first LLM call came back empty before "
+                "any tool ran; retrying",
+                room_id,
+            )
+            try:
+                return await agent.kickoff_async(prompt)
+            except Exception as retry_exc:
+                if _is_empty_llm_response(retry_exc):
+                    logger.warning(
+                        "Room %s: CrewAI's retry also came back empty; giving up",
+                        room_id,
+                    )
+                else:
+                    logger.warning(
+                        "Room %s: CrewAI's retry failed with a different error: %s",
+                        room_id,
+                        retry_exc,
+                    )
+                raise
+
     async def _report_error(self, tools: AgentToolsProtocol, error: str) -> None:
         """Send error event (best effort)."""
         try:
diff --git a/tests/adapters/test_crewai_adapter.py b/tests/adapters/test_crewai_adapter.py
index 92c53642a..48ccae12a 100644
--- a/tests/adapters/test_crewai_adapter.py
+++ b/tests/adapters/test_crewai_adapter.py
@@ -24,6 +24,7 @@
 import pytest
 from pydantic import BaseModel, Field
 
+import band.adapters.crewai as crewai_adapter
 from band.adapters.crewai import EMPTY_LLM_RESPONSE_MARKER
 from band.core.types import Capability, Emit, PlatformMessage
 from band.runtime.prompts import render_system_prompt
@@ -741,6 +742,51 @@ async def test_empty_answer_with_no_tool_call_still_raises(
             )
 
         mock_tools.send_event.assert_awaited_once()
+        # First call plus the one retry -- proves the retry actually happens
+        # before the final raise, not a bare pass-through of the first failure.
+        assert mock_crewai_agent.kickoff_async.call_count == 2
+
+    @pytest.mark.asyncio
+    async def test_empty_first_call_recovers_on_retry(
+        self, CrewAIAdapter, sample_message, mock_tools, mock_crewai_agent
+    ):
+        """A turn's first LLM call coming back empty is retried once before
+        giving up -- a successful retry must complete the turn normally.
+
+        Nothing ran before the empty completion, so there is no side effect
+        for the retry to duplicate; the retry either behaves exactly like the
+        first attempt would have or, as here, recovers and replies.
+        """
+        mock_result = MagicMock()
+        mock_result.raw = "Hello! I'm here to help."
+
+        async def _kickoff(_prompt):
+            if mock_crewai_agent.kickoff_async.call_count == 1:
+                raise ValueError(EMPTY_LLM_RESPONSE_ERROR)
+            tracker = crewai_adapter._reply_tracker_var.get()
+            if tracker is not None:
+                tracker.replied = True
+            return mock_result
+
+        mock_crewai_agent.kickoff_async = AsyncMock(side_effect=_kickoff)
+
+        adapter = CrewAIAdapter()
+        await adapter.on_started("TestBot", "Test bot")
+        adapter._crewai_agent = mock_crewai_agent
+
+        await adapter.on_message(
+            msg=sample_message,
+            tools=mock_tools,
+            history=[],
+            participants_msg=None,
+            contacts_msg=None,
+            is_session_bootstrap=True,
+            room_id="room-123",
+        )
+
+        assert error_events(mock_tools) == []
+        # First call plus the one retry that recovered it.
+        assert mock_crewai_agent.kickoff_async.call_count == 2
 
     @pytest.mark.asyncio
     async def test_no_missing_reply_error_after_clean_tool_only_return(