Agentic AI
Agents
LLM, can think, reflect, use tools, orchestrates the task with long and short term planning
● Agents are software programs designed to autonomously execute
tasks for users or other applications.
● They balance latency and cost to enhance task effectiveness.
● In situations that require greater complexity, workflows provide
reliability and consistency for clearly defined tasks, while agents
excel in scenarios demanding flexibility and scalable, model-driven
decision-making.
● A strong model is necessary for reasoning capabilities.
● Major components
○ LLM
○ Tools / Function calling
○ Memory
Types of Agents
Simple ReAct Agent from Scratch
Build a ReAct Agent
● Agentic AI Fundamentals
○ Building blocks of an AI agent
○ System prompts and message handling
○ Basic conversation flow management
● ReAct Pattern Implementation
○ Thought Action Observation Answer Cycle
○ Step-by-step reasoning handling
○ Function calling and response handling
● Function Integration
○ Defining callable functions (calculate, model_memory)
○ Observation processing and state management
● Practical Application
○ Real-world example: ML model memory calculator
○ Multi-step problem solving
○ Automated reasoning with external tools
Multi Agent System
E-commerce workflow
● Function Calling Architecture
○ LLM agents can invoke real-world functions
○ Pattern: Think Act Observe Respond
○ Structured function definitions with validation
● Multi-Step Reasoning (ReAct Pattern)
○ Agents break complex tasks into steps
○ Each step involves reasoning and action
○ Observations feedback into reasoning
● Workflow Orchestration
○ Managing conversation state and context
○ Coordinating between user input, LLM, and functions
○ Automated action execution and response handling
● System Design Patterns
○ Message-based communication
○ Stateful conversation management
SFT Dataset Generation System
learn to build a sophisticated structured training datasets builder workflow
● System Architecture Design ● Practical Application
○ Dataset creation pipeline architecture ○ Real-world example: SFT dataset creation
○ Content extraction and processing flows ○ Processing different content sources
○ Asynchronous operation handling ○ Quality control and validation
○ Multi-stage data transformation ○ Output file generation and management
● Content Processing Implementation
○ URL and text input handling
○ Content extraction and cleaning
○ Schema validation and enforcement
○ Data format standardization
● Orchestration System
○ Coordinated processing workflows
○ Input validation and error handling
○ Progress tracking and logging
○ State management across stages
Agentic System
Agentic AI Ecosystem
Agentic AI Workflow
● User -> AI Agent: User requests trigger the AI Agent, the central
orchestrator.
● AI Agent -> LLM: The Agent passes requests to the Large
Language Model (LLM) for processing.
● LLM + Databases: LLM uses internal knowledge and retrieves
info from:
○ Database: Structured data
○ Vector Database: Semantic context (e.g., RAG)
● LLM -> Action: LLM directs the AI Agent to perform intelligent
actions.
● Data Flywheel & Customization: Action outcomes generate
data, fueling continuous Model Customization for ongoing
improvement.