Build AI Agents: Homework
This use case explores the foundational concepts behind designing, evaluating, and deploying AI agents in business workflows. The assignment required analyzing how agents process tasks, make decisions, follow constraints, and integrate into real organizational settings. I focused on understanding the difference between simple prompting, system role design, tool-enhanced agents, and multi-step reasoning frameworks.
This homework directly supports the CIDM 6096 objectives related to strategic AI adoption, workflow clarity, and building organizational AI readiness.
Overview of the Assignment
In this homework, I examined:
What makes an AI agent “agentic”
How tasks are decomposed into steps
Where AI should act autonomously vs. be constrained
How guardrails prevent errors
How agents integrate with human workflows
Example tasks that benefit from agent design
Risks, failure modes, and ethical considerations
The focus was on practical business applications, including:
Customer service triage
Document drafting
Research automation
Structured data extraction
Compliance-aligned workflows
Decision support systems
Key Concepts Learned
1. Agents Need Role, Purpose, and Boundaries
An effective AI agent must have:
A clearly defined role
A scope of work
Guardrails and non-permissions
Specific formats and reasoning steps
Without this clarity, agents hallucinate or generalize too broadly.
2. Agentic Behavior Requires Multi-Step Reasoning
A prompt like “Summarize this” is not an agent.
An agent requires:
Goals
Sub-tasks
Step sequences
Validation or checking
Completion criteria
This aligns with modern agent frameworks like AgentKit, LangChain, and Anthropic’s ReAct-style agents.
3. Human Oversight Is Non-Negotiable
The homework emphasized that even well-designed agents:
Can misinterpret ambiguous inputs
Need human review for high-stakes actions
Should default to clarifying questions when uncertain
4. Tool Use Expands Capability (But Adds Risk)
Tool-enabled agents (e.g., API callers, calculators, search tools):
Expand decision-making
Allow complex workflows
Introduce operational risk
Require constraints on when/why tools are used
5. Strong Prompts Strengthen Agent Behavior
I experimented with prompts that included:
Role definition
Step-by-step process instructions
Input/Output requirements
Safety constraints
Examples for pattern learning
These greatly improved stability and reliability.
Example: My Agent Specification Structure
Below is the structure I used to design better agents in this homework:
This structure later informed my custom GPT work (e.g., DME Coverage Decoder).
🤖 Agent Behavior in Action (Evidence)
I tested this structure by simulating a Customer Service Agent.
Prompt: Act as the agent defined above. A customer is asking for a refund outside the 30-day window.
Evidence of Work:
Why This Use Case Matters
This homework strengthened my ability to:
Break down complex workflows
Identify which parts AI can perform reliably
Design agent roles that support real organizational tasks
Anticipate risks, failures, and operational edge cases
Build resources others can use to evaluate agent-based systems
It demonstrates thoughtful, safe, strategic use of GenAI — not just surface-level experimentation.
How This Connects to My Portfolio
The lessons from this assignment show up across multiple areas of my portfolio:
DME Coverage Decoder GPT → Uses strong role, constraints, and output formatting
Workflow Vendor Comparison → Evaluates tools that power agentic workflows
Translation + Patient Education → Demonstrates structured reasoning
TeleCare Triage Thinking → Applies agentic logic to patient intake workflows
This homework forms the backbone of how I now think about building reliable, safe, and useful AI systems.