Use predefined manager_agent

app.py

@@ -3,16 +3,9 @@ import requests
 
 import gradio as gr
 import pandas as pd
-from smolagents import (
-    DuckDuckGoSearchTool,
-    WikipediaSearchTool,
-    VisitWebpageTool,
-    ToolCallingAgent,
-)
 import openlit
 
-from src.models import OllamaModel
-from src.tools_registry import ToolsRegistry
+from src.agents import manager_agent
 
 # (Keep Constants as is)
 # --- Constants ---
@@ -45,17 +38,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
 
     # 1. Instantiate Agent ( modify this part to create your agent)
     try:
-        agent = ToolCallingAgent(
-            tools=[
-                DuckDuckGoSearchTool(),
-                VisitWebpageTool(),
-                WikipediaSearchTool(content_type="text"),
-            ],
-            max_steps=20,
-            model=OllamaModel(model_id="qwen3:14b", tools_registry=ToolsRegistry),
-            planning_interval=5,
-            add_base_tools=True,
-        )
+        agent = manager_agent
     except Exception as e:
         print(f"Error instantiating agent: {e}")
         return f"Error initializing agent: {e}", None

src/agents.py

@@ -0,0 +1,68 @@
+from smolagents import (
+    DuckDuckGoSearchTool,
+    VisitWebpageTool,
+    ToolCallingAgent,
+    CodeAgent,
+)
+
+from src.models import OllamaModel
+from src.tools_registry import ToolsRegistry
+
+
+def check_reasoning_and_answer(final_answer, agent_memory):
+    multimodal_model = OllamaModel(model_id="qwen3:14b", tools_registry=ToolsRegistry)
+    prompt = (
+        "You are a LLM which task is to check the summary of a reasoning and an answer given by other managed LLMs when queried to realise a task."
+        "Here is a user-given task and the agent steps: "
+        f"{agent_memory.get_succinct_steps()}."
+        "Here is the final answer:"
+        f"{final_answer}"
+        "Please check that the reasoning process and the answer are correct: do they correctly answer the given task?"
+        "To help you realize this task, here is some guidances:"
+        "1. First, list reasons why yes/no"
+        "2. Then check the answer is either a string (one or a few words), a number, or a comma separated list of strings or floats, unless specified otherwise. There is only one correct answer.  "
+        "3. Conclude writing your final decision: PASS in caps lock if it is satisfactory, FAIL if it is not."
+        "4. Don't be harsh, if the task couldn't achieve a perfect reasoning because direct data couldn't be found, just PASS in caps lock."
+    )
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "text",
+                    "text": prompt,
+                }
+            ],
+        }
+    ]
+    output = multimodal_model(messages).content
+    print("Feedback: ", output)
+
+    if "FAIL" in output:
+        raise Exception(output)
+
+    return True
+
+
+web_agent = ToolCallingAgent(
+    tools=[
+        DuckDuckGoSearchTool(),
+        VisitWebpageTool(),
+    ],
+    model=OllamaModel(model_id="qwen3:14b", tools_registry=ToolsRegistry),
+    name="web_agent",
+    description="Browses the web to find information",
+    verbosity_level=0,
+    max_steps=10,
+)
+
+
+manager_agent = CodeAgent(
+    model=OllamaModel(model_id="qwen3:14b"),
+    tools=[],
+    managed_agents=[web_agent],
+    planning_interval=5,
+    verbosity_level=2,
+    final_answer_checks=[check_reasoning_and_answer],
+    max_steps=15,
+)

src/models.py

@@ -9,7 +9,7 @@ from ollama import Message as OllamaMessage
 
 from tools_registry import ToolsRegistry
 
-DEFAULT_NUM_CTX = 8192
+DEFAULT_NUM_CTX = 16384
 
 
 class OllamaModel(ApiModel):