Agentic AI guide spans transformer foundations to multi-agent deployment

Haggai Roitman released The Hitchhiker's Guide to Agentic AI on June 25, a full-stack reference for practitioners building autonomous AI systems. The book, available as an arXiv preprint, treats agentic AI as an engineering discipline that requires fluency at every layer—from GPU systems and model compression up through multi-agent coordination protocols.

The first half establishes the substrate: transformer architecture, training methods including supervised fine-tuning and LoRA, mixture-of-experts routing, and inference optimization. Roitman then covers alignment and reasoning—RLHF, PPO, DPO and its variants, GRPO, reward modeling, and reinforcement learning for chain-of-thought and test-time scaling. The second half is devoted to agentic systems proper, with chapters on trajectory-based RL, retrieval-augmented generation (including Agentic RAG), memory architectures, and agent harness design.

What stands out

1. 01Inter-agent coordination gets dedicated coverage. The book devotes multiple chapters to the Model Context Protocol (MCP), the Agent-to-Agent (A2A) communication protocol, and multi-agent topologies—centralized, decentralized, and hierarchical. Most agentic AI writing stops at single-agent tool use; Roitman carries the thread through to production multi-agent systems.
2. 02Memory systems are treated as a first-class design layer. Four memory types—in-context, external, episodic, and semantic—are analyzed with implementation patterns. The book positions memory architecture as a core determinant of agent capability, not an afterthought.
3. 03Production deployment is in scope from the start. The final chapters cover agent development frameworks, agentic UI design, evaluation methodology for agentic tasks, and deployment. Each chapter pairs theory with code examples and references to the primary literature.
4. 04