diff --git a/orchestrator/llm_runner.py b/orchestrator/llm_runner.py
new file mode 100644
index 00000000..527586d9
--- /dev/null
+++ b/orchestrator/llm_runner.py
@@ -0,0 +1,88 @@
+import json
+import logging
+import os
+from datetime import datetime, timezone
+
+from orchestrator.config import OrchestratorConfig
+from orchestrator.signals import Signal
+
+logger = logging.getLogger(__name__)
+
+
+class LLMRunner:
+    def __init__(self, config: OrchestratorConfig):
+        from tradingagents.graph.trading_graph import TradingAgentsGraph
+
+        trading_cfg = config.trading_agents_config if config.trading_agents_config else None
+        self.graph = TradingAgentsGraph(config=trading_cfg)
+        self.cache_dir = config.cache_dir
+        os.makedirs(self.cache_dir, exist_ok=True)
+
+    def get_signal(self, ticker: str, date: str) -> Signal:
+        """获取指定股票在指定日期的 LLM 信号，带缓存。"""
+        cache_path = os.path.join(self.cache_dir, f"{ticker}_{date}.json")
+
+        if os.path.exists(cache_path):
+            logger.info("LLMRunner: cache hit for %s %s", ticker, date)
+            with open(cache_path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            direction, confidence = self._map_rating(data["rating"])
+            return Signal(
+                ticker=ticker,
+                direction=direction,
+                confidence=confidence,
+                source="llm",
+                timestamp=datetime.fromisoformat(data["timestamp"]),
+                metadata=data,
+            )
+
+        try:
+            _final_state, processed_signal = self.graph.propagate(ticker, date)
+            rating = processed_signal if isinstance(processed_signal, str) else str(processed_signal)
+            direction, confidence = self._map_rating(rating)
+            now = datetime.now(timezone.utc)
+
+            cache_data = {
+                "rating": rating,
+                "direction": direction,
+                "confidence": confidence,
+                "timestamp": now.isoformat(),
+                "ticker": ticker,
+                "date": date,
+            }
+            with open(cache_path, "w", encoding="utf-8") as f:
+                json.dump(cache_data, f, ensure_ascii=False, indent=2)
+
+            return Signal(
+                ticker=ticker,
+                direction=direction,
+                confidence=confidence,
+                source="llm",
+                timestamp=now,
+                metadata=cache_data,
+            )
+        except Exception as e:
+            logger.error("LLMRunner: propagate failed for %s %s: %s", ticker, date, e)
+            return Signal(
+                ticker=ticker,
+                direction=0,
+                confidence=0.0,
+                source="llm",
+                timestamp=datetime.now(timezone.utc),
+                metadata={"error": str(e)},
+            )
+
+    def _map_rating(self, rating: str) -> tuple[int, float]:
+        """将 5 级评级映射为 (direction, confidence)。"""
+        mapping = {
+            "BUY": (1, 0.9),
+            "OVERWEIGHT": (1, 0.6),
+            "HOLD": (0, 0.5),
+            "UNDERWEIGHT": (-1, 0.6),
+            "SELL": (-1, 0.9),
+        }
+        result = mapping.get(rating.upper() if rating else "", None)
+        if result is None:
+            logger.warning("LLMRunner: unknown rating %r, falling back to HOLD", rating)
+            return (0, 0.5)
+        return result