The Hitchhiker's Guide to Agentic AI: From Foundations to Systems

The Hitchhiker's Guide to Agentic AI: From Foundations to Systems is a single-author book by Haggai Roitman, posted to arXiv as 2606.24937 on June 22, 2026 and revised on July 27 and September 29, 2026 (the current v3 is labelled version 1.4, CC BY-SA 4.0). A reading group running it chapter by chapter describes it as more than 570 pages. The first half is the substrate under an agent: transformer architecture, GPU systems, training and fine-tuning (SFT, LoRA, mixture of experts), compression and inference optimization, then alignment and reasoning with RLHF, PPO, DPO and its variants, GRPO, reward modelling and test-time scaling. The second half is about agents themselves: trajectory-based agentic RL, RAG and agentic RAG, memory systems, harness design and loop engineering, graph-based orchestration, a catalogue of design patterns that includes red teaming and gateways, inter-agent coordination through MCP, agent skills and the A2A protocol, and centralized, decentralized and hierarchical multi-agent topologies. It closes with agent frameworks, agentic UI, evaluation methodology, production deployment and regulation (the EU AI Act and California SB 942). The MIT-licensed companion repo has one self-contained notebook per chapter. Each notebook adds a piece to the Guardian Angel System, a five-agent capstone for supervising diabetic patients, built with LangChain, LangGraph, CrewAI, AutoGen, Qdrant, Redis and an MCP tool server, running on Ollama with qwen2.5:7b or an OpenAI key.

Topic

Agentic

Level

Intermediate

Format

Book

Cost

Free

Duration

~570-page book, several weeks of reading; 28 companion notebooks plus a capstone

Provider

Haggai Roitman (arXiv)

Hands-on

Yes — code/exercises

Certificate

None

• ✓Frame LLM interaction as a Markov decision process with policies and value functions
• ✓Implement PPO with clipping and a KL penalty, then compare it against DPO and GRPO
• ✓Build a reward model and recognise reward hacking in preference-based training
• ✓Build a RAG pipeline with dense retrieval, re-ranking and Qdrant collections tuned with HNSW and quantisation
• ✓Give agents episodic, semantic and working memory through a shared Redis layer
• ✓Wire stateful supervisor graphs in LangGraph with conditional routing, checkpoints and human approval gates
• ✓Expose agent tools to any model through a Model Context Protocol server
• ✓Evaluate agent answers with LLM-as-judge, G-Eval, RAGAS and FactScore, and red-team them against prompt injection
• ✓Deploy agents as Dockerised FastAPI microservices with health checks, tracing and Redis Streams queues

• •One author covers the whole stack, from GPU systems and RLHF to MCP, A2A and the EU AI Act, so the vocabulary stays consistent across 570+ pages
• •Every chapter notebook adds a part to one running capstone, a five-agent Guardian Angel System, so the code accumulates instead of resetting each chapter
• •The notebooks run locally on Ollama with qwen2.5:7b, fall back to an OpenAI key, and have a mock mode for CI; only the QLoRA chapter needs a GPU
• •Actively revised: three arXiv versions between June and September 2026, a Hugging Face paper page with 22 upvotes, and a community reading group working through it chapter by chapter

This page was written from 5 sources, 4 on domains other than arxiv.org.

1. 1.arxiv.org — 2606.24937vendor
2. 2.alphaxiv.org — 2606.24937
3. 3.huggingface.co — 2606.24937
4. 4.raw.githubusercontent.com — README.md
5. 5.meetup.com — 315595570