Roitman posts a full-stack agentic AI reference to arxiv

There is a certain kind of arxiv upload that tells you more about the state of a field than any single benchmark result would. A comprehensive, single-author guide to agentic AI, posted to arxiv on June 22, 2026 by Haggai Roitman, is one of them. The interesting thing is not any specific claim in the paper. It is that one person can now credibly attempt to map the entire stack, from transformer foundations to production agent deployment, in one document.

The scope is deliberate. The guide walks from the LLM substrate, covering transformer architecture, GPU systems, training and fine-tuning approaches like SFT, LoRA and MoE, model compression, and inference optimization, up through the alignment and reasoning layer, which takes in RLHF, PPO, DPO and its variants, GRPO, reward modeling, and chain-of-thought and test-time scaling for large reasoning models. Then it moves to the agentic layer proper: trajectory-based RL, retrieval-augmented generation and its agentic variant, memory systems split into in-context, external, episodic and semantic, agent harness design, a taxonomy of agent design patterns, and inter-agent coordination via the Model Context Protocol and the Agent-to-Agent (A2A) communication protocol. The closing section covers agent development frameworks, agentic UI design, evaluation methodology for agentic tasks, and production deployment.

Why this matters if you are not the author: a shared reference vocabulary is genuinely useful right now. Practitioners moving into agent work have been reading disjointed blog posts, framework docs, and RL papers. A single guide that names the layers the same way, and pairs each layer with implementation guidance, code examples, and references to the primary literature, lowers the onboarding cost for teams building this stuff for the first time. That is real leverage for hiring managers and training programs.

The honest caveat is what a single-author survey cannot do. It cannot benchmark competing design patterns against each other, and this one does not appear to include quantitative results of its own. The taxonomy is Roitman's, which means readers should cross-check against the primary literature the guide points to rather than treat categories like agent harness, or the MCP versus A2A split, as settled. A Version 1.2.2 reference is a snapshot, and the field is moving fast enough that it will need updates.

The forward-looking value is exactly that vocabulary problem. If bootcamps, corporate training programs, and internal engineering onboarding start pulling from the same stack map, arguing about agentic AI gets a little cheaper for everyone.