agency-agents-zh: 277 role prompts for Chinese-market AI workflows, with caveats
🎭 277 个即插即用的 AI 专家角色 — 支持 Claude Code/Cursor/Copilot 等 20 种工具，覆盖工程/设计/营销/金融等 20 个部门。含 64 个中国市场原创智能体（小红书/抖音/微信/飞书/钉钉/Qt 上位机/机械设计）。搭配编排器 agency-orchestrator，一句话即可让多位专家按 DAG 自动协作。
At a glance
- What is it?
- A Chinese community fork of the agency-agents repository, adding 64 original agents for platforms like Xiaohongshu, Douyin, and WeChat. The value is real for Chinese-language teams, but the README's heavy sponsor content and count inconsistencies deserve scrutiny.
- Who is it for?
- Adopt agency-agents-zh if you work primarily in Chinese and need ready-made role prompts for platforms like Xiaohongshu, Douyin, or WeChat, or if you want a broader set of department roles than the upstream English project. Do not adopt it if you require strict English-only workflows or if you are uncomfortable with a README that mixes sponsor promotions and affiliate links with project documentation.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Shell, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What this fork actually adds
The repository is organized as a set of agent definition files, each containing a system prompt and instructions. The README states that agents are 'plug and play' and support 20 AI coding tools including Claude Code, Cursor, and Copilot. The exact file format is not described in the provided material, but the project's topic list includes 'agent-definitions', 'cursor-rules', and 'system-prompt', which suggests each agent is packaged for import into those tools. The intended user is an engineer or team lead who wants to assemble a virtual company of AI specialists without writing custom prompts from scratch. The companion project, agency-orchestrator, is a desktop client that lets you invoke multiple agents in a DAG-based workflow with a single sentence. The orchestrator is distributed as a native app for macOS, Windows, and Linux, and does not require Node.js. An online demo is available at ao.aiolaola.com/experts.
How the agents are meant to work
The repository's primary language is Shell, which is unusual for a prompt collection. It likely means the project uses shell scripts to install or copy agent files into tool-specific directories. The README does not show any installation commands, so a user must infer the process from the repository structure or the npm package. The npm page is mentioned via a badge, but no package name or installation instructions appear in the README. That is a gap: a project that calls itself 'plug and play' should document how to plug, yet the README is largely a marketing page with sponsor banners and course links. The actual usage instructions are probably in the repository's other files or in the desktop client. For an engineer evaluating this, the first step is to clone the repo and inspect the directory layout, because the README alone does not tell you where agents live or how to load them into Claude Code or Cursor.
Getting started: what the README does and does not tell you
The README is explicit about where to go for a working setup. It points to the agency-orchestrator releases page for a desktop client that runs on macOS, Windows, and Linux without Node.js. It also links to an online experience at ao.aiolaola.com/experts. These are the only concrete 'getting started' paths in the material. There is no command-line installation shown, no git clone example, and no configuration snippet. The npm badge suggests a package exists, but the README does not say how to install it. The release notes mention version v1.2.6 from 2026-06-16, with v1.2.5 titled 'README online display area + simplification' and v1.2.4 titled 'count completion 216 + count guard'. Those titles hint that the repository has undergone reorganization and count checks, but they do not provide usage details. For a technical reviewer, this is a red flag: the project's own documentation is not self-sufficient. You must either use the desktop client or explore the repository files manually to understand the agent format.
A genuine limitation: count inconsistency and documentation quality
The README claims 277 total agents, with 213 translated and 64 original. The release note for v1.2.4 says 'count completion 216 + count guard', which suggests that at some point the total was 216, not 277. The current README says 277, and the release is v1.2.6, so the count may have been corrected, but the history reveals that counts have been a recurring issue. The release note for v1.2.5 mentions 'README online display area + simplification', which may have removed some content. This inconsistency matters because the project's core value proposition is its inventory of agents. If the count is unreliable, you cannot trust the coverage claims. The README also contains a large amount of sponsor content: multiple affiliate links for API resellers and model aggregation services. That is not inherently a problem, but it makes the README feel more like a landing page than technical documentation. The actual agent definitions are not described in any detail: no sample prompt, no directory tree, no format specification. An evaluator cannot determine whether an agent for 'Qt industrial host computer' is a deep specialist or a shallow paragraph without cloning the repo. That is a real limitation for a project that claims to be 'plug and play'.
When this is the wrong tool
If your team works primarily in English, this fork adds little over the upstream agency-agents, and the Chinese-original agents are irrelevant. The translation is for Chinese speakers; English users would be better served by the original repository. Also, if you need agents that integrate with a specific internal tool or API, this project may not fit. The agents are static prompt definitions; they do not include code or connectors. The orchestrator is a separate project, and its capabilities are not documented in this README. If you need a multi-agent system with custom logic, this prompt collection is not a framework. It is a content library. The wrong use case is expecting runtime behavior or state management from the agents themselves. Another wrong use case is compliance-sensitive environments: the README contains promotional claims from sponsors about API reliability and pricing, but that is not project functionality. The agents themselves may produce outputs that need human review, especially in regulated fields like medical compliance or government sales, which the original agents cover. The README does not mention any guardrails or validation mechanisms.
A real alternative: upstream agency-agents
The most direct alternative is the upstream project, agency-agents by msitarzewski. The README links to it and describes this fork as a 'Chinese community edition'. The upstream version contains the original 213 agents that this fork translates. If you do not need Chinese-language definitions or the China-specific additions, the upstream project is the source of truth and likely has more contributors and a longer track record. The difference in approach is that upstream is English-first and does not include the 64 original China-market agents. For a Chinese team, the fork is more relevant. For an international team, upstream is cleaner because it avoids the sponsor-heavy README and the count ambiguity. Another alternative is to write your own role prompts. If you only need a handful of specialists, creating system prompts for Claude Code or Cursor is straightforward. The value of this project is the breadth and the pre-packaged nature, not the technical complexity. So the choice is between breadth and control.
Maintenance, licensing, and what to verify before adoption
The project is licensed under MIT, which permits commercial use, modification, and redistribution with attribution. That is a permissive license, but do not treat this as legal advice. The repository is not archived, and the last push was on 2026-09-09, with releases up to v1.2.6 in June 2026. That indicates active maintenance as of the material date. However, the README contains many sponsor banners and affiliate links, which suggests the project is monetized through those partnerships. That is not a technical problem, but it means the README is not a neutral documentation source. The release notes show a pattern of count corrections, so before adopting, verify the actual number of agent files in the repository. Also check whether the agents are compatible with your specific tool version, since the README lists 20 tools but does not specify which agents work with which tool. The companion orchestrator is a separate project with its own release cycle; evaluate it separately. If you plan to rely on the online demo, note that it is a hosted service, not a self-hosted solution. For production use, you would likely clone the repo and load agents into your own toolchain.
Editorial conclusion
Adopt agency-agents-zh if you work primarily in Chinese and need ready-made role prompts for platforms like Xiaohongshu, Douyin, or WeChat, or if you want a broader set of department roles than the upstream English project. Do not adopt it if you require strict English-only workflows or if you are uncomfortable with a README that mixes sponsor promotions and affiliate links with project documentation. Before using it, verify the actual agent count in your cloned copy: the README claims 277 total and 64 original, but release notes mention a 'counting fix to 216', which suggests past discrepancies. Also check which of the 20 listed tools have actual definitions in the repository, since the README does not provide a per-tool breakdown.
Community notes