AG-2 in Practice #5 – Building Custom Patterns and Integrating Tools

#ag2 #ai

Hey again!

So far, you’ve seen AG-2’s built-in patterns and multi-agent setups. Now, it’s time to level up by creating your own custom pattern with conditional logic — plus, we’ll integrate a tool so agents can perform actions beyond chatting.

Why Custom Patterns?

Sometimes, you want workflows that:

Retry if quality checks fail
Branch based on agent feedback
Use tools dynamically during conversations

Custom patterns give you that full control.

Step 1: Define a Custom Pattern with Conditional Logic

Here’s an example pattern where:

Agent A writes a draft
Agent B reviews it
If review fails, Agent A rewrites and resubmits

from ag2 import Agent, Pattern, Conversation
import os

os.environ["OPENAI_API_KEY"] = "your-api-key-here"

class ReviewRewritePattern(Pattern):
    def __init__(self, writer, reviewer, max_retries=2):
        super().__init__()
        self.writer = writer
        self.reviewer = reviewer
        self.max_retries = max_retries

    def run(self, prompt):
        retries = 0
        while retries <= self.max_retries:
            # Step 1: Writer drafts
            draft = self.writer.chat(prompt)

            # Step 2: Reviewer reviews
            review = self.reviewer.chat(draft)

            # Step 3: Check if approved
            if "approve" in review.lower():
                return draft

            retries += 1
            prompt = f"Rewrite based on review: {review}"

        return f"Max retries reached. Last draft:\n{draft}"

# Agents
writer = Agent(
    name="writer",
    llm="openai/gpt-4",
    system_message="Write clear and detailed responses.",
)

reviewer = Agent(
    name="reviewer",
    llm="openai/gpt-4",
    system_message="Review the text carefully and say 'approve' if good, otherwise give feedback.",
)

pattern = ReviewRewritePattern(writer, reviewer)
conv = Conversation(pattern=pattern)
result = conv.send("Explain quantum computing in simple terms.")
print(result)

Step 2: Adding a Tool to Your Agent

Let’s add a simple math tool that your writer agent can call to do calculations.

from ag2 import Tool

def calculator(a: int, b: int) -> int:
    return a + b

calc_tool = Tool(
    name="calculator",
    description="Performs addition of two integers",
    func=calculator,
)

writer.tools = [calc_tool]

Now your writer can call calculator in its reasoning.

Step 3: Run the Full Workflow

python custom_pattern_tool.py

Watch the interaction as the writer drafts, the reviewer approves or requests edits, and the writer uses the calculator if needed.

What’s Next?

In upcoming lessons, we’ll explore:

More complex tool integrations (APIs, databases)
Human-in-the-loop workflows for final approval
Debugging and monitoring agent conversations

Keep coding

DEV Community