🧠 AI Agent Workshops

13 Feb 2025, University of Wollongong, GiftCity, Gandhinagar

---

📌 Table of Contents

🔹Part 1: Building Blocks – The Augmented LLM

• Basic LLM Calls – Sending requests and processing responses
• Structured Output – Formatting responses in structured formats
• Tool Use – Integrating external tools with the LLM
• Retrieval – Using memory and external sources for better responses

🔹Part 2: Workflow Patterns for AI Systems

• Prompt Chaining – Structuring multi-step AI tasks
• Routing – Directing requests to specialized handlers
• Parallelization – Running multiple AI processes simultaneously

🔹Part 3: Introduction to Agentic Framework (CrewAI)

• CrewAI Framework Step by Step
• Agent : Finance Agent App

🔹Part 4: Introduction to Agentic Framework(SmolAgent)

• SmolAgent Framework Step by Step
• Agent : AI News Agent App

Sr No	Topic	Colab Link
1.	CrewAI Notebook to Make A Finance Agent
2.	SmolAgent Streamlit App Run in Notebook(AI news Agent)
3.	CrewAI Gemini Streamlit App Run in Notebook(AI News Agent)
4.	Gemini Vision Pro Code Agent(Code Agent)

---

🔄Part 2: Workflow Patterns for AI Systems

🏗️ Prompt Chaining

Prompt chaining breaks down complex AI tasks into smaller, more manageable steps. Each step processes and validates the output from the previous step, improving control and reliability.

📅 Calendar Assistant Example

graph LR
    A[User Input] --> B[LLM 1: Extract]
    B --> C{Gate Check}
    C -->|Pass| D[LLM 2: Parse Details]
    C -->|Fail| E[Exit]
    D --> F[LLM 3: Generate Confirmation]
    F --> G[Final Output]

Loading

📝 Step Breakdown:

Step	Description
✅ Step 1: Extract & Validate	- Detects if input is a validcalendar request - Assigns a confidence score - Filters out irrelevant requests
✅ Step 2: Parse Details	- Extracts structured information (date, time, participants) - Converts natural language into a structured format
✅ Step 3: Generate Confirmation	- Creates auser-friendly response - Generates calendar links if necessary

---

🔀 Routing

Routing is a pattern that directs different types of requests to specialized handlers, allowing for clean separation of concerns and optimized processing.

📅 Calendar Assistant Example

graph LR
    A[User Input] --> B[LLM Router]
    B --> C{Route}
    C -->|New Event| D[New Event Handler]
    C -->|Modify Event| E[Modify Event Handler]
    C -->|Other| F[Exit]
    D --> G[Response]
    E --> G

Loading

📝 Step Breakdown:

Component	Description
✅ Router	- Classifies requests intonew event or modification - Provides confidence scoring
✅ Specialized Handlers	-New Event Handler: Creates calendar events - Modify Event Handler: Updates existing events

---

⚡ Parallelization

Parallelization improves efficiency by running multiple LLM calls simultaneously to analyze different aspects of a request in parallel.

📅 Calendar Assistant Example

graph LR
    A[User Input] --> B[Calendar Check]
    A --> C[Security Check]
    B --> D{Aggregate}
    C --> D
    D -->|Valid| E[Continue]
    D -->|Invalid| F[Exit]

Loading

📝 Step Breakdown:

Component	Description
✅ Parallel Checks	- Calendar Validation: Ensures a valid request - Security Check : Screens for prompt injection
✅ Aggregation Layer	- Merges results from parallel checks - Makes thefinal validation decision

---

🎭 Orchestrator-Workers

The orchestrator-workers pattern uses a central LLM to dynamically analyze, coordinate, and synthesize responses from specialized workers. This is useful for tasks requiring structured content generation.

📝 Blog Writing Example

graph LR
    A[Topic Input] --> B[Orchestrator]
    B --> C[Planning Phase]
    C --> D[Writing Phase]
    D --> E[Review Phase]
    style D fill:#f9f,stroke:#333,stroke-width:2px

Loading

📝 Step Breakdown:

🛠️Orchestrator	📝Planning Phase	✍️Writing Phase	🔍Review Phase
🔎 Analyzes theblog topic and requirements	📌 Breaks content intosections	🏗️ Assigns sections tospecialized writers	✅ Evaluatescontent flow and cohesion
🏗️ Generates astructured content plan	📏 Definesword count and writing style	🔗 Maintainscontext and consistency	✨ Suggestsimprovements
🔄 Overseescontent cohesion	-	-	🏆 Produces apolished final version

---

🔄Part 3: Introduction to Agentic Framework 🤖

This section introduces additional architectures for building AI agents, providing a structured overview of their workflows and modular designs.

CrewAIAgent

CrewAI follows a modular, step-by-step approach that includes:

Example: Finance Agent App

Workflow

graph TD
    A[Start] --> B[Initialize Streamlit App]
    B --> C[Set Up API Keys & Load Environment Variables]
    C --> D[User Inputs Company Name]
    D --> E{Start Analysis Button Clicked?}
  
    E -- Yes --> F[Display Progress & Setup Agents]
    F --> G[Create Financial Analyst Agent]
    G --> H[Create Investment Strategy Reviewer Agent]
    H --> I[Initialize Stock Market Scraper Tool]
    I --> J[Define Stock Market Analysis Task]
    J --> K[Define Investment Review Task]
    K --> L[Create Financial Analysis Crew]
  
    L --> M[Run AI Analysis Crew Process]
    M --> N[Generate Financial Report]
    N --> O[Display Final Financial Analysis Report]
  
    E -- No --> P[Wait for User Action]

    style A fill:#ffcc00,stroke:#333,stroke-width:2px
    style O fill:#ffcc00,stroke:#333,stroke-width:2px
    style M fill:#00ccff,stroke:#333,stroke-width:2px
    style J fill:#00ccff,stroke:#333,stroke-width:2px

Loading

🔄Part 4: Introduction to Agentic Framework(SmolAgent)

SmalAgent

SmolAgent provides a live coding guide for building lightweight agents. Key highlights include:

• Streamlit App: Running using Colab as the backend server
• AI News Agent App

Workflow

graph TD
    A[Start] --> B[Initialize Streamlit App]
    B --> C[Initialize LLM and Search Tool]
    C --> D[User Inputs News Topic]
    D --> E[Set Search Depth and Analysis Type]
    E --> F{Analyze News Button Clicked?}
  
    F -- Yes --> G[Perform DuckDuckGo Search]
    G --> H{Results Found?}
  
    H -- Yes --> I[Create Analysis Prompt]
    I --> J[Generate Analysis Using LLM]
    J --> K[Display Analysis Results]
    K --> L[Log Activity]
  
    H -- No --> M[Show No Results/Error Message]
  
    F -- No --> N[Wait for User Action]
  
    K --> O[Show Tips for Better Results]
    M --> O
    O --> P[End]

    style A fill:#ffcc00,stroke:#333,stroke-width:2px
    style P fill:#ffcc00,stroke:#333,stroke-width:2px
    style G fill:#00ccff,stroke:#333,stroke-width:2px
    style J fill:#00ccff,stroke:#333,stroke-width:2px

Loading

Running live in Colab

graph TD
    SA[SmolAgent]
    SA --> SC[Step-by-Step Live Coding]
    SC --> S[Streamlit App via Colab]
    SC --> N[AI News Agent App]

Loading

---

🤝 Contributing

Contributions are welcome! To get started:

1. Fork the repository
2. Create a branch (feature-new-pattern)
3. Commit your changes
4. Push to GitHub and open a PR

For detailed guidelines, check CONTRIBUTING.md.

---

📜 License

This project is licensed under the MIT License – see LICENSE for details.

---

📬 Contact

For questions or collaborations, feel free to reach out:

📧 Email: ashishpatel.ce.2011@gmail.com

---

🎯 Happy Coding! 🚀