The Hitchhiker’s Guide to Agentic AI

A book-length practitioner’s reference by Haggai Roitman, posted to arXiv (cs.AI) on 22 June 2026 and revised 27 July 2026. Its thesis is a claim about competence rather than about architecture: building good agentic systems requires understanding every layer of the pipeline, not one of them. The structure follows that argument top to bottom.

• LLM substrate — transformer architecture, GPU systems, SFT, LoRA, MoE, model compression, inference optimization. Explicitly framed as foundation, not subject.
• Alignment and reasoning — RLHF, PPO, DPO and variants, GRPO, reward modeling, RL for large reasoning models, chain-of-thought and test-time scaling.
• Agentic systems — trajectory-based RL for agent training, RAG and Agentic RAG, memory (in-context, external, episodic, semantic), harness design and context management, loop engineering (inference-time RL, generate-verify-retry, adaptive budget control), and a taxonomy of agent design patterns.
• Coordination — MCP, agent skills and tool use, the A2A protocol, and multi-agent topologies: centralized, decentralized, hierarchical.
• Delivery — agent development frameworks, agentic UI, evaluation methodology for agentic tasks, production deployment.

Every chapter pairs theory with implementation guidance, code and citations to the primary literature.

T1 — a named, revised, citable work by a single author with its sources exposed. What it is not is new evidence: it is a synthesis of the literature, so its claims are as good as the papers underneath them, and a reader should follow the references rather than cite the guide.

Why this one matters to the spoke

The corpus has been built the opposite way — one repository, one blog post, one benchmark at a time, with synthesis doing the assembling. This is the first source that arrives already assembled, and it names as chapters several things this wiki had to derive from scattered tools.

loop-engineering is the clearest case. That page was built up from practitioner accounts (autoresearch‘s overnight training loop, the Ralph technique, verification threads) and argues the loop is the design unit and verification the real bottleneck. Roitman gives the same idea a chapter with named mechanisms — inference-time RL, generate-verify-retry optimization, adaptive budget control — and the last of those is a lever nothing in the corpus has documented. Same for agent-memory: the spoke reached the in-context/external/episodic/semantic split empirically through iai-pme‘s three tiers and jcode‘s memory graph; here it is the standing taxonomy.

The reverse is also worth stating. The book covers evaluation methodology for agentic tasks as a chapter, and evaluation is exactly where this corpus keeps finding nothing — cc-thinking-skills is still the only pack anyone ablated, and the day’s three big packages (superclaude-framework, ruflo, opc-skills) shipped none between them. A reference that treats agent evaluation as settled enough to teach, over a field that does not measure its own artifacts, is the tension to hold rather than resolve.

Cross-spoke context. The substrate and alignment halves are ../machine-learning-wiki’s and ../llm-inference-wiki’s subjects, and the whole thing is also a curriculum artifact of the kind ../engineering-education-wiki studies. Routed here because the second half — harnesses, loops, memory, MCP, A2A, multi-agent topologies — is this spoke’s domain and is what the book is for.

Related

loop-engineering · agent-memory · agentic-coding-harness · agent-orchestration · agent-evals · agentskills-spec · model-context-protocol · spec-driven-development · context-rot · synthesis