02-Running-Agents.md

🚀 02 Running Agents

Create a script in the project root of your repository e.g shopping_agents.py and follow these steps:

1. Add your imports:

import asyncio
import os
from dotenv import load_dotenv
from autogen_agentchat.agents import UserProxyAgent
from autogen_agentchat.conditions import TextMentionTermination
from autogen_agentchat.teams import RoundRobinGroupChat
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import AzureOpenAIChatCompletionClient
from autogen_ext.agents.web_surfer import MultimodalWebSurfer

2. Load your envrionment variables from your .env file:

load_dotenv()

3. Create variables from your .env file:

api_key = os.getenv("AZURE_OPENAI_API_KEY")
azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT")
api_version = os.getenv("AZURE_OPENAI_API_VERSION")

4. Define your model:

model_client = AzureOpenAIChatCompletionClient(
     model="gpt-4o",
     api_key=api_key,
     azure_endpoint=azure_endpoint,
     azure_deployment="gpt-4o",
     api_version=api_version
)

5. Define your agents:

#Web Surfer Agent
web_surfer = MultimodalWebSurfer(
     name="web_surfer",
     model_client=model_client,
     headless=False,           # Enables visible browser window
     animate_actions=True
 )

#User Proxy Agent
user_proxy = UserProxyAgent(name="user_proxy")

6. Create your team with the RoundRobinGroup Chat and set Termination condition:

team = RoundRobinGroupChat(
    participants=[web_surfer, user_proxy],
    termination_condition=TextMentionTermination("exit", sources=["user_proxy"])
)

The team uses RoundRobinGroupChat to alternate turns between the agents. The session ends when the user_proxy agent receives the word "exit" from your terminal input.

7. Define your main function with stream logic:

async def main():
    # Start the task and stream responses to the terminal
    stream = team.run_stream(
        task="Browse the e-commerce site https://www.currys.co.uk/ and add headphones to the shopping basket."
    )
    
    # The Console streams the agent interactions to your terminal live
    await Console(stream)

    # Close connections after execution
    await web_surfer.close()
    await model_client.close()

8. Run the main function:

if __name__ == "__main__":
    asyncio.run(main())
    print("🛍️ Shopping Complete: ✅")

🧩 How It All Comes Together

Once you have configured your agents, run your script in the terminal:

python shopping_agents.py

A browser window should pop up, watch how the agent is navigating the web on its own. This process is fully autonomous - you don’t need to interact with the browser. Just observe as the agent explores the page. You may see a red dot moving around and clicking around the webpage, like a human would!

In the terminal, the Console class prints real-time updates of each agent's decisions and tool use, giving full observability into the task. If you toggle between the web browser and the terminal you can see a full break down of what the agents are doing:

1. The task instruction tells the team what to do (e.g. “Go to an e-commerce site and add headphones to the cart”)
2. The agents take turns thinking and acting
3. The WebSurfer opens the browser and performs the web navigation
4. The UserProxyAgent lets you observe or interact if needed in the terminal
5. When done, you type exit to stop in the terminal

---

⚠️ There may be times when agents fail to complete a task. Common reasons include:

• Security restrictions (e.g CAPTCHA, login prompts, bot detection)
• Incomplete instructions or vague goals
• Unsupported site layouts or dynamic JavaScript rendering

---

Designing reliable multi-agent systems is a real-world engineering challenge. Coordination often breaks down so building effective agentic workflows is an iterative process of trial, observation, and refinement.

💡 Creative Challenges

These extra tasks help you explore AutoGen’s flexibility and push the limits of your web agent:

💭 Task 1:

• Did your agent manage to add items to the basket successfully?
  If not, examine the terminal logs to understand what went wrong. Try:

  • Rephrasing the task more specifically (e.g "add the first item to the basket")

  • Asking the assistant to redirect to an alternative site

  • Adding intermediate steps (e.g “search first, then filter by price”)

    For example:

      stream = team.run_stream(
        task="Browse the e-commerce site https://www.amazon.co.uk/ and add headphones to the shopping basket. Add the first item to the basket."
    )

💭 Task 2:

• Use the UserProxyAgent to interact mid-run:
  Type guidance, clarifications, or commands directly into the terminal - you’re part of the loop!

💭 Task 3:

• Create an additional custom AssistantAgent with a unique system message. For example, add an AssistantAgent with this behavior: “Instruct the WebSurfer agent to try alternative product pages if a site returns no results.”

  Try replacing the UserProxyAgent entirely with your custom AssistantAgent, to enable a more autonomous experience.

  You can read more about the AssistantAgent and different types of other agents here: 🔗 AutoGen Agent Guide

⚠️ Don't forget to add any new agents you define to your group chat!

⚠️ Don't forget to update your termination condition (or comment it out) if you are no longer using the UserProxyAgent for termination.

💭 Task 4:

• Try another group chat type like:

  • SelectorGroupChat: Automatically routes queries to the most relevant agent.

  You can read more about the different types of group chats here: 🔗AutoGen AgentChat Guide

You will find SelectorGroupChat under the Advanced section along with other types of chat

Have fun exploring AutoGen! It’s a powerful platform for building smart, collaborative agents that can read, search, click, and reason across the web!🌐