New 'Hitchhiker's Guide' covers full stack of agentic AI from LLMs to multi-agent systems

Haggai Roitman has released 'The Hitchhiker's Guide to Agentic AI' (arXiv:2606.24937), a 500+ page practitioner's reference that systematically covers the entire stack of autonomous AI systems. The book opens with what Roitman calls the "LLM substrate" — transformer architecture, GPU systems, training techniques (SFT, LoRA, Mixture of Experts), model compression, and inference optimization — establishing these as essential foundations rather than the main focus. It then moves into alignment and reasoning, providing deep coverage of RLHF, PPO, DPO and its variants, GRPO, reward modeling, and reinforcement learning for large reasoning models including chain-of-thought and test-time scaling. Each topic is explained with both theoretical rigor and practical implementation guidance.

The second half of the book is devoted entirely to agentic AI. Chapters cover agentic training and trajectory-based RL, retrieval-augmented generation (both standard RAG and Agentic RAG), and a taxonomy of memory systems (in-context, external, episodic, semantic). Roitman then dives into agent harness design, context management, and a catalog of agent design patterns. Inter-agent coordination gets thorough treatment, including the Model Context Protocol (MCP), tool use, the Agent-to-Agent (A2A) communication protocol, and multi-agent architectures (centralized, decentralized, hierarchical). The book concludes with frameworks for development, agentic UI design, evaluation methodology for agentic tasks, and guidance on production deployment. With code examples and references to primary literature throughout, it positions itself as a one-stop reference for engineers building production-ready autonomous AI.