# Hermes Agent Orange Book: what the 2.0 PDF covers and where it is wrong > The Orange Book is a free PDF guide to Nous Research's Hermes Agent, rebuilt around v0.16.0. It is a book about a framework, not the framework, and it ships with a self-reported erratum. **alchaincyf/hermes-agent-orange-book** — Hermes Agent 从入门到精通 · 橙皮书系列 · Nous Research 开源 AI Agent 框架实战指南 - Repository: https://github.com/alchaincyf/hermes-agent-orange-book - Stars: 4,957 · Forks: 493 - Language: Python - License: NOASSERTION - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/alchaincyf-hermes-agent-orange-book ## A book about Hermes Agent, not the agent itself The repository alchaincyf/hermes-agent-orange-book contains no runnable code for the framework. The top level holds two PDFs, a LICENSE, a README in English and Chinese, image assets and a tools directory. The subject of the book is a separate project, Nous Research's Hermes Agent, hosted at github.com/NousResearch/hermes-agent. If you arrived looking for the agent's source, you are in the wrong repository. The audience the README names is narrow and specific. Developers who have used Claude Code, OpenClaw or Cursor and want to understand how Hermes differs. AI power users who avoid the command line, since the framework now ships a desktop app. And readers who want to see Harness Engineering concepts turned into a shipping product. The README frames Hermes as the productization of five components: instructions, constraints, feedback, memory and orchestration. That framing is the book's actual thesis. It is written by HuaShu, an AI content creator and indie developer, as part of a free 橙皮书 series that also covers Claude Code, Harness Engineering and OpenClaw. Treat it as one practitioner's structured notes on a fast-moving framework, not as vendor documentation. ## What the 2.0 rewrite changed and why it matters The first edition was built on Hermes v0.7.0. The README says the second edition is a from-scratch rewrite around v0.16.0, which it calls "The Surface Release", after nine releases and roughly two months. The structure went from 17 chapters in 5 parts to 21 sections in 6 parts. The new material is concentrated in three areas the first edition barely touched: the self-improvement loop and its Curator component, the multi-agent Kanban platform, and an OS-level security model. The README also lists a persistent multi-agent Kanban platform and automatic Skill creation and evolution as things the framework grew in this cycle. Section 4 of the book covers 64 tools, MCP, 23 messaging platforms and three surfaces. A rewrite at this cadence is a signal about the framework, not the book. Nine releases in two months means any guide is a snapshot with a short shelf life. The README acknowledges this directly, telling readers to refer to the official docs for the latest. That is an honest admission, and it also tells you what the book is for: orientation and mental models, not current API surface. ## The erratum is the most useful page in the repository The README devotes a full section to an error in section 03, "Why Nous Built It". The book claimed Hermes Agent doubles as a data harvester, drawing a chain from user usage to tool-call trajectories to training data for Nous's own models. The author states that chain breaks at its first link. Two sources are cited. The official FAQ says API calls go only to the LLM provider you configure, that Hermes Agent does not collect telemetry, usage data or analytics, and that conversations, memory and skills are stored locally in ~/.hermes/. The project's AGENTS.md lists "outbound telemetry / usage attribution without opt-in gating" as a contribution that gets rejected even when well built. The author's own explanation of the mistake is the interesting part: the artifacts the book cited, batch_runner.py and trajectory_compressor.py, are real, but they are research infrastructure for batch-generating trajectories in your own environment, not a pipeline collecting data from users. He stretched "they have trajectory-generating infrastructure" into "they are harvesting your trajectories". The README also notes the original text flagged this as inference rather than an official statement, and argues that a flag does not rescue an inference the official docs contradict. Two consequences follow. The "data flywheel" argument at the end of section 03 is withdrawn, while the other attributed motives, monetizing through Portal and competing for OpenClaw's users, still stand. And the PDFs are build artifacts that have not been regenerated, so the published files still contain the uncorrected text. You are reading a correction that has not reached the artifact you download. ## Getting the PDF and checking whether it is current There is nothing to install. The README gives a download table with two entries, one Chinese and one English, both pointing at files on the main branch. The English edition is fetched with a plain HTTP download: ```bash curl -L -o Hermes-Agent-The-Complete-Guide-v260607.pdf \ https://github.com/alchaincyf/hermes-agent-orange-book/raw/main/Hermes-Agent-The-Complete-Guide-v260607.pdf ``` You should end up with a PDF named Hermes-Agent-The-Complete-Guide-v260607.pdf in your working directory. The filename encodes the build date, v260607, which is the marker to look at. If a later build appears on main with a different suffix, that one may carry the corrected section 03. To check the repository state without downloading anything, clone it shallowly and look at the top level: ```bash git clone --depth 1 https://github.com/alchaincyf/hermes-agent-orange-book.git ls hermes-agent-orange-book ``` The listing should show the two PDFs, LICENSE, README.md, README_zh.md, assets/, banner.png, cover.png, screenshots/ and tools/. If the PDF filenames still end in v260607, the erratum has not been folded into the body text yet. ## Where the book stops being the right tool The book is pinned to Hermes Agent v0.16.0. The README describes nine releases between the first and second editions, which is roughly one every week or so. Anything version-sensitive in the text, a flag, a config key, a tool count, is a candidate for having changed. The README does not document a changelog per section or mark which passages are version-bound, so you cannot tell from the outside which paragraphs are fragile. The erratum exposes a second problem. The book mixes documented product facts with the author's own inference, and at least once the inference was wrong in a way the official docs already contradicted. The author flags inferences, which helps, but a flag is not a verification step. If you need to know what the framework actually does with your data, read the official FAQ and AGENTS.md rather than the chapter that summarizes them. Finally, no releases are listed for this repository, and the last push was on 2026-08-23. The PDFs are build artifacts that the README says have not been regenerated, so the repository can receive commits that do not change the file you downloaded. A recent commit date is not evidence that the PDF content is current. There is also a language trap in the search results. People search for "Agent Orange" books, which is a herbicide and a historical subject, not this framework. The Orange Book name here comes from the 橙皮书 series branding. Nothing in this repository is about that other topic. ## Alternatives: official docs, or the Harness Engineering volume The obvious alternative is the official documentation at hermes-agent.nousresearch.com/docs, which the README itself points to for product facts. The difference in approach is structural. The docs are maintained by the project and track the current release; the book is maintained by one author and pinned to a version. The docs will tell you the current flag name. The book will tell you why the framework has a Curator, a three-layer memory system and a Kanban surface, and how those pieces relate. Those are different jobs and neither substitutes for the other. A second alternative sits inside the same series. The README says Hermes is the productization of the five components described in the Harness Engineering 橙皮书: instructions, constraints, feedback, memory and orchestration. If your interest is the design pattern rather than this specific framework, that volume is the more direct route, and it will not go stale every time Hermes ships a release. If you want to evaluate the framework itself, neither book helps. The repository has no dependency on Hermes Agent, no pinned version, no test harness. The only thing you can verify from here is whether the prose matches the docs. ## Licence and what upgrading costs you The README states that starting with 2.0 the book is under MIT. The first edition used CC BY-NC-SA 4.0. The repository metadata reports the licence as NOASSERTION, which means GitHub could not classify the LICENSE file automatically. The README's statement and the repository metadata do not agree on their face, and resolving that is a job for whoever needs certainty, not a conclusion this article can draw. The practical difference between the two licences is the non-commercial clause. CC BY-NC-SA 4.0 restricts commercial use and requires share-alike. MIT does not. If you plan to reuse chapters in training material or an internal handbook, the README says MIT applies from 2.0 onward, and the LICENSE file in the repository is the thing to read before you rely on that. Upgrade cost is the harder question. There is no migration path between editions, because 2.0 is a rewrite rather than a patch. The first edition, v260408, is not listed in the download table, so readers who need v0.7.0 coverage have to find it through the repository history. And because the PDFs are build artifacts that have not been regenerated, the erratum in the README is currently the only corrected copy of section 03. Budget for reading two documents, not one. ## Conclusion Read the Orange Book if you already run Hermes Agent and want a structured walkthrough of the self-improvement loop, the memory layers and the multi-agent Kanban surface, or if you are deciding whether the framework is worth an afternoon. Do not read it as product documentation: the README itself says to trust the official docs for product facts, and the PDFs on main have not been regenerated since the erratum was posted. Before you rely on any chapter, check the erratum in the README first, then confirm the section you care about against the official docs, because the book is pinned to v0.16.0 and the framework has moved since. ## FAQ ### How much does the Hermes Agent Orange Book cost? Nothing. The README lists both the Chinese and English PDFs as free downloads, and the series page describes all books in the 橙皮书 series as free to download. ### Are there books about Agent Orange? This repository is not one of them. The Orange Book name here comes from the 橙皮书 series branding used by this Hermes Agent guide, and nothing in the repository concerns the herbicide or its history. ### Is there a Hermes Agent dashboard? The README says the framework grew a full browser dashboard alongside a native desktop app, and that the book's part 4 covers three surfaces. It does not describe the dashboard's features, so check the official docs for that. ### Who is the Hermes Agent Orange Book written for? The README names developers who have used Claude Code, OpenClaw or Cursor, AI power users who avoid the command line, and readers interested in Harness Engineering concepts turned into a product. ## Sources - [alchaincyf/hermes-agent-orange-book on GitHub](https://github.com/alchaincyf/hermes-agent-orange-book) - [Issues](https://github.com/alchaincyf/hermes-agent-orange-book/issues) - [README](https://github.com/alchaincyf/hermes-agent-orange-book/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/alchaincyf-hermes-agent-orange-book