diff --git a/skyvern/forge/agent.py b/skyvern/forge/agent.py
index d960c756..0e350719 100644
--- a/skyvern/forge/agent.py
+++ b/skyvern/forge/agent.py
@@ -1487,7 +1487,11 @@ class ForgeAgent:
                     reason=json_response.get("thought"), error_type=json_response.get("error")
                 )
 
-            action_type: str = json_response.get("action_type") or ""
+            inferred_actions: list[dict[str, Any]] = json_response.get("inferred_actions", [])
+            if not inferred_actions:
+                raise FailedToParseActionInstruction(reason=json_response.get("thought"), error_type="EMPTY_ACTION")
+
+            action_type: str = inferred_actions[0].get("action_type") or ""
             action_type = ActionType[action_type.upper()]
 
             if action_type == ActionType.CLICK:
diff --git a/skyvern/forge/prompts/skyvern/infer-action-type.j2 b/skyvern/forge/prompts/skyvern/infer-action-type.j2
index cc343ac1..f018899b 100644
--- a/skyvern/forge/prompts/skyvern/infer-action-type.j2
+++ b/skyvern/forge/prompts/skyvern/infer-action-type.j2
@@ -4,10 +4,14 @@ MAKE SURE YOU OUTPUT VALID JSON. No text before or after JSON, no trailing comma
 
 Reply in the following JSON format:
 {
-    "thought": str, // A string to describe how to infer the action type from the action instruction.
-    "confidence_float": float, // The confidence of the action. Pick a number between 0.0 and 1.0. 0.0 means no confidence, 1.0 means full confidence
-    "action_type": str, // It's a string enum: "CLICK", "INPUT_TEXT", "UPLOAD_FILE", "SELECT_OPTION". "CLICK" means user wants to click. "INPUT_TEXT" means user wants to input text. "UPLOAD_FILE" means user wants to upload a file. "SELECT_OPTION" means user wants to select an option.
-    "error": str, // It's a string enum to describe error. Null if you can identify the action as one of the defined action type. Use "UNKNOWN_ACTION" if none of the defined action type matched. Use "MULTIPLE_ACTIONS" if the instruction includes multiple actions.
+    "thought": str, // A string to describe how to infer the action types from the action instruction.
+    "inferred_actions": array // An array of action types. The result should be sorted in descending order by confidence_float. If there are multiple actions, each action should have an unique confidence_float value.
+    [{
+        "reasoning": str, // A string to explain the reason for inferring the action from the instruction.
+        "confidence_float": float, // The confidence of the action. Pick a number between 0.0 and 1.0. 0.0 means no confidence, 1.0 means full confidence.
+        "action_type": str, // It's a string enum: "CLICK", "INPUT_TEXT", "UPLOAD_FILE", "SELECT_OPTION". "CLICK" means user wants to click. "INPUT_TEXT" means user wants to input text. "UPLOAD_FILE" means user wants to upload a file. "SELECT_OPTION" means user wants to select an option. 
+    }],
+    "error": str, // It's a string enum to describe error. Null if you can identify any defined action from the instruction. Use "UNKNOWN_ACTION" if none of the defined action type matched.
 }
 
 Action instruction