The Rise of the Agent

A "Chatbot" is reactive: it answers questions based on a fixed context. An "Agent" is proactive: it can determine which steps are needed to accomplish a goal, call external tools to gather data, and course-correct when it hits an error.

In the AWS Certified Generative AI Developer – Professional exam, Domain 2 places a massive focus on Agentic AI Patterns. You must demonstrate that you can design the "Reasoning Loop" that allows an agent to solve problems like a human developer.

---

1. Agent vs. Chatbot: The Key Differences

Feature	Chatbot	AI Agent
Logic	Fixed prompting.	Dynamic planning.
Tools	None (usually just RAG).	Can call APIs, run code, and search the web.
Autonomy	Follows a strict script.	Self-directs based on the goal.
Memory	Simple history.	Task management and state tracking.

---

2. The REACT Pattern (Reason + Act)

Most modern agents follow the ReAct (Reasoning and Acting) framework. Instead of jumping straight to an answer, the model follows a cycle:

1. Thought: "To answer this, I first need to check the inventory database."
2. Action: Call Tool: search_inventory(item="shoes")
3. Observation: "The database says we have 0 blue shoes in stock."
4. Reflection/Thought: "Since blue is out, I should check for black shoes before responding."

graph TD
    User[User Goal] --> P[Plan Step 1]
    P --> T[Execute Tool/Action]
    T --> O[Observe Result]
    O --> R{Goal Met?}
    R -->|No| P2[Plan Step 2]
    P2 --> T
    R -->|Yes| End[Final Answer]

---

3. Introducing Amazon Bedrock Agents

Writing a ReAct loop from scratch in Python is difficult because you have to handle "State" and "Infinite Loops." Amazon Bedrock Agents is a fully managed service that does the heavy lifting for you.

Core Components:

• Foundation Model: The "Brain" (e.g., Claude 3.5).
• Instructions: The "Personality" and "Boundary" (e.g., "You are a travel assistant for internal employees only").
• Action Groups: The "Hands" (Lambda functions or API schemas).
• Knowledge Bases: The "External Memory" (RAG).

---

4. Activity Planning and Decomposition

A Professional Agent doesn't just "guess" the next step. It performs Task Decomposition.

Scenario: User says "Book a trip to London for next Tuesday and find a hotel under $200." The Agent's Plan:

1. Check flight availability on Tuesday.
2. Confirm price vs budget.
3. Search for London hotels.
4. Filter by price (<$200).
5. Present options to the user.

---

5. Avoiding the "Infinite Loop"

One major risk with agents is that they get stuck in an endless loop (Thought -> Action -> Thought -> Action). Pro Developer Strategy:

• Set a Max Iterations limit (e.g., 5-10 loops max).
• Use Stop Sequences in your prompts.
• Implement Human-in-the-loop for critical or repeated failures.

---

6. Code Fragment: Defining an Agent Instance

import boto3

bedrock_agent = boto3.client('bedrock-agent')

def create_travel_agent():
    response = bedrock_agent.create_agent(
        agentName='InternalTravelAssistant',
        foundationModel='anthropic.claude-3-sonnet-20240229-v1:0',
        instruction='You help employees book flights and hotels within policy.',
        agentResourceRoleArn='arn:aws:iam::123456789012:role/bedrock-agent-role'
    )
    return response['agent']['agentId']

---

Knowledge Check: Test Your Agent Knowledge

---

Summary

Agents are the future of GenAI. By giving a model "Instructions" and "Tools," you turn it into a digital employee. In the next lesson, we will look at exactly how these "Hands" work: Tool Integration and Stateful Decision Systems.

---

Next Lesson: The Hands of AI: Tool Integration and Stateful Decision Systems