With the advances in computer vision, speech recognition, and large language models, we can now equip computers with perception and generative ability. Large language models brained Agents in particular have been a key breakthrough, allowing Agentic AI to reason its world through natural language and leverage broad knowledge to make smarter decisions.
This course aims to explore the theoretical foundations of Agentic AI and the platforms and tools for developing AI agents. The focus will be guiding students in creating autonomous agents capable of operating independently, making decisions in complex environments, and completing sophisticated tasks.
Learning Objectives
Upon completing this course, students will be able to:
Understand the fundamental principles and architecture of Agentic AI systems.
Design and implement essential AI agents using modern frameworks and tools.
Consider ethical and safety aspects in AI agent design.
Evaluate the performance and limitations of Agentic AI systems.
Assess the challenges AI agents face when interacting in real-world environments.
Prerequisites
Programming background (Python)
Familiarity with basic artificial intelligence concepts
Fundamental knowledge of machine learning and neural networks
This hands-on course focuses on developing autonomous AI agents with practical applications, emphasizing web interaction, multi-agent systems, memory management, and human-in-the-loop learning. Students will progressively build an agent system through four interconnected assignments and a final project.
Weeks 1: Course Introduction
● Introduction to agency in AI systems
● Historical development of autonomous agents
● Agent architectures and frameworks
● Core components: perception, reasoning, and action
Weeks 2-4: Assignment 1 - Reflexive Web Agent with Tools Use (15%)
● Language models brained agents
● Prompt engineering for agency
● Tool use and API interaction
● LLM Reasoning: Chain-of-Thoughts, Tree-of-Thoughts
● Schema-guided task-oriented dialog systems
Technical Focus
● Prompt Engineering: Assign specialized roles and pre-built and custom tools to agents
● Basic web interaction patterns
● Selenium
Weeks 5-7: Assignment 2 - Self-Reflection via MultiAgent Collaboration (15%)
● Agent communication and coordination framework: Collaborative and competitive agents
● Decision-making under uncertainty
● Self-Reflection mechanism
● Evaluation of task result for decision-making
Technical Focus
● Break down the tasks, goals, and tools and assign to multiple AI agents for better performance
● LangChain, LangGraph, AutoGen, CrewAI
Weeks 8-10: Assignment 3 - Agents with Memory Recall via RAG (15%)
● Memory management systems: Provide agents with short-term, long-term, and shared memory
● Goal representation and task decomposition
● Information retrieval based on embedding
● Retrieval augmented generation
● Object reference resolution
Technical Focus
● RAG architecture design
Week 11: Midterm Examination (15%)
Weeks 12-14: Human in the Loop Reinforcement Learning (15%)
● Experience recording
● Memory persistence
● Task decomposition and Hierarchical architectures
● Adaptive behavior patterns
Technical Focus
● Ollma
Weeks 15-18: Final Project (25%)
● Integrate all previous assignments to a real-world environment
● Define environment and actions, e.g. LINE APP, Windows OS, or Smart Home APP
● Guardrails: Effectively handle errors, hallucinations, and infinite loops
3 Programming Assignments: 45%
Paper Reading & Oral Report: 10%
Involvement & Discussion: 10%
Mid-term Exam: 15%
Final Project: 20%