Agentic AI Bootcamp

A practical, intensive 3-day hands-on bootcamp covering Agentic AI that takes you from first principles to a fully working, self-built autonomous AI agent.

You will learn core agentic AI concepts and architectures, work with leading open-source tooling, host and fine-tune models locally, equip agents with memory and custom skills, orchestrate multi-agent systems, and apply essential security defenses — including protection against prompt injection.

By the end of the bootcamp, you will have designed, trained, secured, and demonstrated your own locally hosted agentic assistant.

What You’ll Build in 3 Days

• Day 1: Your first ReAct agent — a tool-using Q&A agent powered by an open-source model, complete with a visible reasoning trace.
• Day 2: A locally hosted open-source model (via Ollama), custom skills/tools for your agent, a RAG knowledge skill using a vector database, and a small model fine-tuned with LoRA on a custom dataset.
• Day 3: A two-agent pipeline (planner + worker), an indirect prompt-injection attack on your own agent, and a practical defense against it.
• Capstone: A complete, locally-hosted agentic assistant that combines custom skills, memory/RAG, and at least one prompt-injection defense — presented live to the group.

Tools & Platforms You’ll Master

• Local Model Hosting: Ollama (pulling models, Modelfiles, quantization)
• Hugging Face Ecosystem: Model Hub, Transformers, PEFT / TRL
• Agent Frameworks: LangChain, LangGraph, CrewAI, AutoGen
• Memory & RAG: ChromaDB (vector memory)
• UI Prototyping: Gradio / Streamlit
• Fine-tuning Techniques: LoRA, QLoRA, instruction datasets
• Open-source Models: Llama, Mistral, Qwen, Phi and other agent-friendly models

Day-by-Day Curriculum

DAY 1: Foundations of Agentic AI

Understanding what agentic AI is, how it differs from a chatbot, and building your first agent.

Topics Covered

• What is Agentic AI — autonomy, tool use, goal-directed reasoning
• Agentic AI vs. traditional chatbots / LLM apps
• Types of Agentic AI: reflex, model-based, goal/planning-based, utility-based, learning agents
• ReAct agents (Reason + Act) and reflection agents
• Single-agent vs. multi-agent systems — overview
• Anatomy of an agent: LLM, memory, tools, planner/orchestrator
• Introduction to agent frameworks: LangChain, LangGraph
• Overview of open-source models suited for agentic tasks (Llama, Mistral, Qwen, Phi)

Hands-on Demo: Build your first ReAct agent — a tool-using Q&A agent powered by an open-source model, with a visible reasoning trace.

Challenges Discussed

• Hallucinated or incorrect tool calls
• Infinite reasoning loops and latency/cost of multi-step agents

DAY 2: Building, Training & Hosting Your Agent

Hosting open-source models locally, giving your agent skills, and fine-tuning your own model.

Topics Covered

• Hosting open-source LLMs locally with Ollama — pulling models, Modelfiles, quantization
• The Hugging Face ecosystem — Model Hub, Transformers, choosing a model for agentic use
• Giving your agent skills — custom tool/function creation, plugin-style skill registries
• Retrieval-Augmented Generation (RAG) as a knowledge skill using a vector database
• Training / fine-tuning your agent’s model — when to fine-tune vs. prompt vs. RAG
• Fine-tuning fundamentals: instruction datasets, LoRA and QLoRA, PEFT
• Exporting and hosting a fine-tuned model locally

Hands-on Demo: Host an open-source model locally via Ollama, add custom skills to your agent, and fine-tune a small model with LoRA on a custom dataset.

Challenges Discussed

• GPU/VRAM limits and quantization trade-offs
• Fine-tuning dataset quality and catastrophic forgetting

DAY 3: Multi-Agent Systems & Agentic AI Security

Orchestrating multiple agents, understanding prompt injection, and securing your agent — capped by a capstone build.

Topics Covered

• Multi-agent architectures: supervisor–worker, planner–executor, debate/critic patterns
• Agent memory at scale — short-term vs. long-term, memory drift and hallucination
• Prompt injection fundamentals — direct vs. indirect injection, tool-poisoning, excessive agency
• Mapping to the OWASP Top 10 for LLM Applications
• Defenses and guardrails — least-privilege tool scoping, human-in-the-loop, input/output filtering, sandboxing
• Deploying and demoing a finished agentic application

Hands-on Demo: Build a two-agent pipeline (planner + worker), then attack your own agent with an indirect prompt-injection payload and apply a defense against it.

Challenges Discussed

• Agent hijacking via indirect prompt injection
• Balancing autonomy with safety and control

Capstone Project

• A locally-hosted agentic assistant combining custom skills, memory/RAG, and at least one prompt-injection defense — presented live to the group