Agentic AI and Agent Orchestration

Title: Agentic AI and Agent Orchestration

Tutorial Time: 13.30-16.30 (3 April)

Presenters: Pakdee Jaisue

  • Onsite Registration Fee: 7,500 Bath
  • Online Registration Fee: 5,000 Bath

Abstract:

This three-hour intensive tutorial is designed for developers, AI practitioners, and tech enthusiasts who want to move beyond simple "prompt-and-response" interactions with Large Language Models (LLMs). We will explore the transition from passive LLMs to Agentic AI—systems capable of reasoning, using tools, and executing multi-step tasks autonomously.

Participants will learn the core architecture of an AI agent, how they maintain state and memory, and the strategies used for Agent Orchestration to manage complex workflows involving multiple specialized agents. By the end of this session, you will understand how to design and structure autonomous systems that can solve real-world problems with minimal human intervention.

Learning Objectives:

  • Differentiate between a standard LLM application and an Agentic system.
  • Identify the appropriate orchestration pattern for a given business problem.
  • Design a multi-agent workflow that includes specialized roles and feedback loops.

Syllabus (3-Hour Tutorial):

    Module 1: From LLMs to Agents (45 Minutes)
  • The LLM Foundation: A brief technical refresher on how LLMs process information and the limitations of "stateless" chat.
  • Defining Agentic AI: What makes a system "agentic"? (Reasoning, Planning, and Action).
  • The Agent Loop: Understanding the Perception-Reasoning-Action cycle.
  • Tool Use (Function Calling): How agents interact with the outside world (APIs, databases, and web searching).
    Module 2: Core Components of an Agent (45 Minutes)
  • Planning Strategies: Chain-of-Thought (CoT) and Self-Reflection.
  • Task Decomposition: Breaking complex goals into manageable steps.
  • Memory Management: Short-term memory (Context windows).
  • Long-term memory (Vector databases and RAG integration).
  • Persona and Instructions: Crafting robust system prompts to define agent behavior and constraints.
    Module 3: Agent Orchestration & Multi-Agent Systems (60 Minutes)
  • Why Orchestrate? When one agent isn't enough: handling complexity through specialization.
  • Orchestration Patterns:
    • Router Pattern: Directing tasks to specific expert agents.
    • Sequential Chains: Step-by-step handoffs between agents.
    • Joint Collaboration: Multiple agents working in a shared environment (e.g., Coder, Critic, and Manager).
  • Human-in-the-Loop (HITL): Implementing checkpoints for human approval and guidance.

    • Module 4: Design Patterns & Future Trends (30 Minutes)
  • Reliability & Error Handling: Dealing with hallucinations, infinite loops, and tool failures.
  • Evaluation: How to measure the performance of an autonomous agent.
  • The Road Ahead: The shift toward "Small Language Models" (SLMs) for edge agents and the ethics of autonomy.

    • Prerequisites
  • Basic understanding of Python or general programming logic.
  • Familiarity with Large Language Models via the HuggingFace library.