Giving the AI a Grip

In the previous lesson, we designed the "Brain" of the agent. Now, we must give it "Hands." In Generative AI, these hands are called Tools (or Action Groups in the context of Amazon Bedrock). A tool allows the model to reach outside its neural network and interact with the real world—calculating math, search a database, or sending an email.

In this lesson, we will master Tool Definition, Function Calling, and State Management.

---

1. What is a Tool? (Function Calling)

A tool is simply a piece of code (often a Lambda function) that the model can choose to execute. The model doesn't "run" the function itself; instead:

1. The model says: "I want to run get_weather for city 'London'."
2. Your application (or AWS Bedrock) executes the code and gets the result.
3. Your application sends the result back to the model.

---

2. Defining Tools with JSON Schema

To use a tool, the model must understand the "Interface." We define this using OpenAPI or JSON Schema.

{
  "name": "get_stock_price",
  "description": "Get the current price for a specific stock ticker.",
  "parameters": {
    "ticker": {
      "type": "string",
      "description": "The stock symbol (e.g. AMZN, AAPL)."
    }
  },
  "required": ["ticker"]
}

Pro Tip: The description is the most important part. If you write a bad description, the model won't know when to use the tool.

---

3. Bedrock Action Groups

In Amazon Bedrock Agents, tools are organized into Action Groups.

• The Trigger: You provide an OpenAPI (Swagger) schema.
• The Executor: You provide an AWS Lambda function.
• The Handshake: When the model wants to act, Bedrock automatically reformats the request into a JSON payload and invokes your Lambda.

graph LR
    A[Bedrock Agent] -->|Tool Selection| B[Bedrock Action Group]
    B -->|JSON Payload| C[AWS Lambda]
    C -->|API Call| D[External Service: e.g. Salesforce]
    D -->|Data| C
    C -->|Response| B
    B -->|Observation| A

---

4. Stateful Decision Systems

A "Stateful" agent remembers what happened in previous turns.

• Session State: Using sessionId in Bedrock ensures the model carries the context of the conversation.
• Variables: You can pass "Session Attributes" from your code to the agent. For example, you can tell the agent user_tier = "Platinum" so it knows which tools are allowed for this specific user.

---

5. Handling Tool Failures

What if the database is down? As a Professional Developer, you must handle the "Observation" of a failure.

• The Observation: Instead of crashing, your code should return: "Error: The database reached a timeout. Please try again in 5 minutes."
• The Model's Action: The model will read this error and say to the user: "I'm sorry, I'm having trouble reaching our inventory records right now. Would you like me to try a different task?"

---

6. Code Example: A Pro-Lambda for Action Groups

def lambda_handler(event, context):
    # Extracts parameters passed by the Bedrock Agent
    action_group = event['actionGroup']
    function_name = event['function']
    parameters = event['parameters'] # e.g. [{'name': 'ticker', 'value': 'AMZN'}]
    
    if function_name == 'get_stock_price':
        ticker = parameters[0]['value']
        # Mock API logic
        price = 180.00 
        
        response_body = {
            "TEXT": f"The price of {ticker} is ${price}."
        }
    
    # Format required by Bedrock Action Groups
    response = {
        'actionGroup': action_group,
        'function': function_name,
        'functionResponse': {
            'responseBody': response_body
        }
    }
    return response

---

Knowledge Check: Test Your Tool Knowledge

---

Summary

Tools turn an AI from a "Thinker" into a "Doer." By mastering Action Groups and Lambda integration, you can build agents that actually run your business processes. In the final lesson of Domain 2, we will look at Monitoring Agent Execution.

---

Next Lesson: The Guardian: Monitoring Agent Execution