diff --git a/.gitignore b/.gitignore
index 3fc66cecb..6e1fc7f74 100644
--- a/.gitignore
+++ b/.gitignore
@@ -29,6 +29,7 @@ share/python-wheels/
 MANIFEST
 metagpt/tools/schemas/
 examples/data/search_kb/*.json
+expo/AutogluonModels
 
 # PyInstaller
 #  Usually these files are written by a python scripts from a template
diff --git a/expo/experimenter/autogluon.py b/expo/experimenter/autogluon.py
index 4f5d151ef..478ecfc01 100644
--- a/expo/experimenter/autogluon.py
+++ b/expo/experimenter/autogluon.py
@@ -1,27 +1,29 @@
+from datetime import datetime
 from autogluon.tabular import TabularDataset, TabularPredictor
-
 from expo.experimenter.custom import CustomExperimenter
 
 
 class AGRunner:
     preset = "best_quality"
-    time_limit = 500
+    time_limit = 1000 # 1000s
 
-    def __init__(self, datasets):
-        self.datasets = datasets
+    def __init__(self, state=None):
+        self.state = state
+        self.datasets = self.state["datasets_dir"]
 
     def run(self):
         train_path = self.datasets["train"]
-        test_wo_target_path = self.datasets["test_wo_target"]
         dev_wo_target_path = self.datasets["dev_wo_target"]
+        test_wo_target_path = self.datasets["test_wo_target"]
         target_col = self.state["dataset_config"]["target_col"]
         train_data = TabularDataset(train_path)
-        test_data = TabularDataset(test_wo_target_path)
         dev_data = TabularDataset(dev_wo_target_path)
-
-        predictor = TabularPredictor(label=target_col).fit(train_data, presets=self.preset, time_limit=self.time_limit)
-        test_preds = predictor.predict(test_data)
+        test_data = TabularDataset(test_wo_target_path)
+        eval_metric = self.state["dataset_config"]["metric"].replace(" ", "_")
+        # predictor = TabularPredictor(label=target_col, eval_metric=eval_metric, path="AutogluonModels/ag-{}-{}".format(self.state['task'], datetime.now().strftime("%y%m%d_%H%M"))).fit(train_data, presets=self.preset, time_limit=self.time_limit, fit_weighted_ensemble=False, num_gpus=1)
+        predictor = TabularPredictor(label=target_col, eval_metric=eval_metric, path="AutogluonModels/ag-{}-{}".format(self.state['task'], datetime.now().strftime("%y%m%d_%H%M"))).fit(train_data, num_gpus=1)
         dev_preds = predictor.predict(dev_data)
+        test_preds = predictor.predict(test_data)
         return {"test_preds": test_preds, "dev_preds": dev_preds}
 
 
@@ -30,4 +32,16 @@ class GluonExperimenter(CustomExperimenter):
 
     def __init__(self, args, **kwargs):
         super().__init__(args, **kwargs)
-        self.framework = AGRunner(self.datasets)
+        self.framework = AGRunner(self.state)
+
+    async def run_experiment(self):
+        result = self.framework.run()
+        user_requirement = self.state["requirement"]
+        dev_preds = result["dev_preds"]
+        test_preds = result["test_preds"]
+        score_dict = {
+            "dev_score": self.evaluate_predictions(dev_preds, "dev"),
+            "test_score": self.evaluate_predictions(test_preds, "test"),
+        }
+        results = [0, {"score_dict": score_dict, "user_requirement": user_requirement, "args": vars(self.args)}]
+        self.save_result(results)
\ No newline at end of file
diff --git a/expo/experimenter/custom.py b/expo/experimenter/custom.py
index 4a5486af0..df090fb58 100644
--- a/expo/experimenter/custom.py
+++ b/expo/experimenter/custom.py
@@ -12,7 +12,7 @@ class CustomExperimenter(Experimenter):
 
     def __init__(self, args, **kwargs):
         super().__init__(args, **kwargs)
-        self.framework = kwargs["framework"]  # todo
+        self.framework = kwargs.get("framework", None)  # todo
         self.task = kwargs.get("task", self.args.task)
         self.low_is_better = kwargs.get("low_is_better", self.args.low_is_better)
         self.name = kwargs.get("name", "")
diff --git a/expo/insights/solution_designer.py b/expo/insights/solution_designer.py
index fc05afeea..b1fcf4188 100644
--- a/expo/insights/solution_designer.py
+++ b/expo/insights/solution_designer.py
@@ -19,7 +19,8 @@ DATASET_INSIGHT_PROMPT = """
 Propose insights to help improve the performance of the model on this dataset.
 The insights should be proposed based on the dataset description with different task types.
 Each task type should have at least 5 insights.
-Make sure each method is independent and can be implemented separately.
+Make sure each method is diverse enough and can be implemented separately.
+Be specific about models' choices, ensemble and tuning techniques, and preprocessing & feature engineering techniques.
 
 # Format
 ```json
diff --git a/expo/run_experiment.py b/expo/run_experiment.py
index b68607d79..2123fade3 100644
--- a/expo/run_experiment.py
+++ b/expo/run_experiment.py
@@ -2,6 +2,7 @@ import argparse
 import asyncio
 
 from expo.experimenter.aug import AugExperimenter
+from expo.experimenter.autogluon import GluonExperimenter
 from expo.experimenter.custom import CustomExperimenter
 from expo.experimenter.experimenter import Experimenter
 from expo.experimenter.mcts import MCTSExperimenter
@@ -11,7 +12,10 @@ def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument("--name", type=str, default="")
     parser.add_argument(
-        "--exp_mode", type=str, default="mcts", choices=["mcts", "aug", "base", "custom", "greedy", "random"]
+        "--exp_mode",
+        type=str,
+        default="mcts",
+        choices=["mcts", "aug", "base", "custom", "greedy", "autogluon", "random"],
     )
     get_di_args(parser)
     get_mcts_args(parser)
@@ -51,6 +55,8 @@ async def main(args):
         experimenter = AugExperimenter(args)
     elif args.exp_mode == "base":
         experimenter = Experimenter(args)
+    elif args.exp_mode == "autogluon":
+        experimenter = GluonExperimenter(args)
     elif args.exp_mode == "custom":
         experimenter = CustomExperimenter(args)
     else: