diff --git a/ai-hub/integration_tests/test_browser_llm.py b/ai-hub/integration_tests/test_browser_llm.py
index 608a36c..56c920b 100644
--- a/ai-hub/integration_tests/test_browser_llm.py
+++ b/ai-hub/integration_tests/test_browser_llm.py
@@ -31,10 +31,10 @@
         
         session_id = r_sess.json()["id"]
         
-        # Step 2: Ask a question that requires real-time knowledge (weather)
-        # We explicitly ask it to use its search/browser tools to ensure it invokes them.
+        # Step 2: Ask a question that requires the browser to take a snapshot/screenshot
+        # We explicitly ask it to navigate and snapshot to guarantee image output.
         chat_payload = {
-            "prompt": "Use your browser or search tools to find the current weather in Tokyo, Japan. Briefly summarize it in 1 sentence.",
+            "prompt": "Use your browser_automation_agent tool to navigate to https://example.com, take a screenshot of the page, and tell me the heading you see on the page.",
             "provider_name": "gemini"
         }
         
@@ -57,8 +57,8 @@
                         
                         if event_type == "content":
                             full_response += event.get("content", "")
-                        elif event_type == "status":
-                            if "Dispatching" in event.get("content", ""):
+                        elif event_type == "tool_start":
+                            if event.get("name") == "browser_automation_agent":
                                 tool_invoked = True
                         elif event_type == "error":
                             pytest.fail(f"LLM backend emitted an error: {event.get('content')}")
@@ -67,7 +67,44 @@
                     except json.JSONDecodeError:
                         pass
         
-        full_response = full_response.strip()
+        full_response = full_response.strip().lower()
         assert len(full_response) > 0, "LLM returned an entirely silent response."
-        assert tool_invoked, "The LLM didn't attempt to invoke any web/browser tools as instructed."
-        # If it invoked the tool and formulated a response, the end-to-end framework succeeded!
+        assert tool_invoked, "The LLM didn't attempt to invoke the browser tool as instructed."
+        
+        # The prompt asked for example.com heading ("Example Domain").
+        assert "example domain" in full_response, f"LLM did not identify the correct heading. Response: {full_response}"
+
+        # Step 3: Verify the browser agent physically saved the screenshot and metadata to the file sync system.
+        # Since the session_id is either used directly or via sync_workspace_id wrapper, we search the mirrors directory.
+        import subprocess
+        cmd = [
+            "docker", "exec", "ai_hub_service", "bash", "-c", 
+            "find /app/data/mirrors/ -name '.browser_data' -type d"
+        ]
+        res = subprocess.run(cmd, capture_output=True, text=True)
+        assert res.returncode == 0, f"Failed to search for .browser_data inside Docker: {res.stderr}"
+        
+        # We expect at least one .browser_data directory corresponding to our session
+        found_dirs = res.stdout.strip().split('\n')
+        # match session id or sync_workspace_id (e.g. 'session-21-' or '21')
+        matched_dir = next((d for d in found_dirs if str(session_id) in d), None)
+        assert matched_dir, f"Could not find .browser_data directory for session {session_id} in mirrors. Found: {res.stdout}"
+        
+        # Now verify there's a screenshot inside it
+        verify_cmd = [
+            "docker", "exec", "ai_hub_service", "bash", "-c",
+            f"find {matched_dir} -name '*.png' -type f | wc -l"
+        ]
+        res_png = subprocess.run(verify_cmd, capture_output=True, text=True)
+        png_count = int(res_png.stdout.strip() or "0")
+        assert png_count > 0, "No screenshot .png files were saved within the browser agent's output folder."
+
+        # Verify metadata / A11y summary saved
+        verify_meta_cmd = [
+            "docker", "exec", "ai_hub_service", "bash", "-c",
+            f"find {matched_dir}/.metadata -name '*.txt' -o -name '*.json' | wc -l"
+        ]
+        res_meta = subprocess.run(verify_meta_cmd, capture_output=True, text=True)
+        meta_count = int(res_meta.stdout.strip() or "0")
+        assert meta_count > 0, "No extraction metadata or A11y text files found in the .browser_data/.metadata folder."
+