Module 10 Wrap-up: The Professional Architect

You have moved beyond the "Wild West" of scripts. You now understand that scale requires Configuration, Lifecycle Management, and Global Policies. For our final exercise of this module, we will design an agent that follows a strict security policy.

---

Hands-on Exercise: The Privacy-First Support Agent

The Goal: Build an agent that reads a user's question, but a global "Guardrail" blocks it if it tries to output an email address or a credit card number.

1. The Configuration (YAML)

agent_type: "CustomerSupport"
model: "gpt-4o"
persona:
  role: "Support Technician"
  backstory: "You help users with software bugs. Be polite and helpful."
policies:
  - name: "PII_REDACTOR"
    type: "output_filter"
    pattern: "email_regex"

2. The Policy Code (Conceptual)

import re

def pii_redactor_policy(output: str):
    # This is a 'Hard Guardrail'
    email_pattern = r'[a-z0-9\.\-+_]+@[a-z0-9\.\-+_]+\.[a-z]+'
    if re.search(email_pattern, output):
        # We REJECT the whole answer rather than just redacting it
        # This forces the agent to realize it violated privacy rules
        raise SecurityViolation("Agent tried to leak an email address.")
    return output

# In your ADK Wrapper
class ADKAgent:
    def run(self, prompt):
        raw_result = self.model.think(prompt)
        
        # APPLY GLOBAL POLICIES AUTOMATICALLY
        for policy in self.global_policies:
            try:
                raw_result = policy.verify(raw_result)
            except SecurityViolation:
                return "I apologize, but I cannot provide that information for security reasons."
        
        return raw_result

---

Module 10 Summary

• ADK standardizes how agents are built, deployed, and managed.
• YAML Configuration makes agents modular and portable.
• Lifecycle Management ensures agents don't "rot" after deployment.
• Global Policies provide safety guardrails that individual developers can't accidentally disable.

Coming Up Next...

In Module 11, we start the Framework Showdown. We will compare LangChain, CrewAI, and LangGraph side-by-side to help you decide which one to use for your next production project.

---

Module 10 Checklist

• I can explain the benefit of using YAML over hard-coded prompts.
• I understand the 5 stages of an agent's lifecycle.
• I can define a "Hard Guardrail" vs. a "Soft Prompt."
• I understand how to propagate a user's token to an enterprise tool.
• I can describe why "Shadow Deployment" (Champion/Challenger) is used.