diff --git a/rdagent/log/timer.py b/rdagent/log/timer.py
index 62716734..bcf252e5 100644
--- a/rdagent/log/timer.py
+++ b/rdagent/log/timer.py
@@ -75,10 +75,12 @@ class RDAgentTimer:
 class RDAgentTimerWrapper(SingletonBaseClass):
     def __init__(self) -> None:
         self.timer: RDAgentTimer = RDAgentTimer()
+        self.api_fail_count: int = 0
+        self.latest_api_fail_time: datetime | None = None
 
     def replace_timer(self, timer: RDAgentTimer) -> None:
         self.timer = timer
         logger.info("Timer replaced successfully.")
 
 
-RD_Agent_TIMER_wrapper = RDAgentTimerWrapper()
+RD_Agent_TIMER_wrapper: RDAgentTimerWrapper = RDAgentTimerWrapper()
diff --git a/rdagent/oai/backend/base.py b/rdagent/oai/backend/base.py
index f00a568e..357f0f86 100644
--- a/rdagent/oai/backend/base.py
+++ b/rdagent/oai/backend/base.py
@@ -11,6 +11,7 @@ from datetime import datetime
 from pathlib import Path
 from typing import Any, Optional, cast
 
+import pytz
 from pydantic import TypeAdapter
 
 from rdagent.core.utils import LLM_CACHE_SEED_GEN, SingletonBaseClass
@@ -349,6 +350,8 @@ class APIBackend(ABC):
                         content[: len(content) // 2] for content in kwargs.get("input_content_list", [])
                     ]
                 else:
+                    RD_Agent_TIMER_wrapper.api_fail_count += 1
+                    RD_Agent_TIMER_wrapper.latest_api_fail_time = datetime.now(pytz.timezone("Asia/Shanghai"))
                     if (
                         openai_imported
                         and isinstance(e, openai.APITimeoutError)
diff --git a/rdagent/utils/workflow.py b/rdagent/utils/workflow.py
index dd9b243f..721ba9e0 100644
--- a/rdagent/utils/workflow.py
+++ b/rdagent/utils/workflow.py
@@ -138,6 +138,20 @@ class LoopBase:
                         + current_local_datetime.year * 1e10
                     )
                     mlflow.log_metric("current_datetime", float_like_datetime)
+                    mlflow.log_metric("api_fail_count", RD_Agent_TIMER_wrapper.api_fail_count)
+                    lastest_api_fail_time = RD_Agent_TIMER_wrapper.latest_api_fail_time
+                    if lastest_api_fail_time is not None:
+                        mlflow.log_metric(
+                            "api_fail_count",
+                            (
+                                lastest_api_fail_time.second
+                                + lastest_api_fail_time.minute * 1e2
+                                + lastest_api_fail_time.hour * 1e4
+                                + lastest_api_fail_time.day * 1e6
+                                + lastest_api_fail_time.month * 1e8
+                                + lastest_api_fail_time.year * 1e10
+                            ),
+                        )
 
                 if self.timer.started:
                     if RD_AGENT_SETTINGS.enable_mlflow: