A new arXiv preprint doesn't report a breakthrough - it tries to write the textbook on agentic AI before anyone else does.
Posted October 12, 2026 as arXiv:2610.13455 in the cs.AI category, the paper is framed less as a traditional research paper and more as a book-length survey. It walks through how agentic large language models work, how they take action in the world, and how they interact with each other and with human users. The abstract notes that LLM output already shapes real decisions, and asks what happens once models can act directly rather than just advise. The listing is marked as a new submission rather than a revision, and the abstract itself cites no experiments, benchmarks, or new model results - this is a conceptual framing document, not a capability announcement.
Agentic frameworks - autonomous coding assistants, multi-step browser agents, the bot swarms that scheduling and research startups have shipped since 2023 - have mostly grown through trial and error, with formal theory playing catch-up. A shared vocabulary for comparing agent architectures, and for reasoning about what happens when models negotiate with each other instead of just answering a prompt, is genuinely missing from the field, which is what makes a synthesis like this one useful.
Whether a single arXiv preprint can serve as that reference text is another question - "book" is an ambitious label for a submission whose own abstract runs one paragraph and names no case studies.