It is a localized expert library, not an AI workforce in a box
agency-agents-zh takes the role-library idea from msitarzewski/agency-agents and turns it into a Chinese-first collection. The headline number is 277 specialists across 20 departments, from executive roles such as CEO and CFO to engineering, design, marketing, security, finance, GIS, and game work. Each entry is presented as more than a generic prompt: the README says roles have their own personas, professional processes, and expected deliverables. That is useful structure, not proof that every role will make sound decisions.
Localization is the reason to choose this edition. The project reports 213 translated upstream roles and 64 original roles for the Chinese market. Those additions cover Xiaohongshu, Douyin, WeChat, Bilibili, Feishu, DingTalk, cross-border commerce, government work, medical compliance, industrial Qt software, mechanical design, and livestock record checking. English readers can use the linked upstream documentation, but this repository is Chinese-first. Teams serving those markets get vocabulary and workflows that an English-only catalog would require them to recreate.
What happened when we ran it
We cloned commit e00aed9 into an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. The npm install completed in 4 seconds, installed 0 packages, and occupied 11 MB. That fits the repository's nature: it is primarily a collection of files and instructions, not an application with a dependency tree. The checkout contained 366 files, about 2,645 lines of source, and occupied 6.2 MB before installation.
There was no build script or target, so our run skipped compilation. There was also no test script or tests directory, so we could not execute an automated quality check. npm audit reported 0 known vulnerabilities across all severity levels, though 0 packages were installed. We found 4 CI workflow files. Those are a positive maintenance signal, but they do not substitute for tests that verify role formatting, installation paths, or behavior across the advertised 20 tools.
The breadth is useful when you know the job
The strongest feature is selection. A team can start with a named specialist rather than stretching one general assistant across 20 departments. The China-specific set is particularly concrete: platform operations, regulated sectors, industrial software, and vertical business tasks are not merely translated labels. The repository also claims compatibility with Claude Code, Cursor, Copilot, and 17 other AI coding tools, making the role files more portable than a catalog locked to one interface. The MIT license allows internal adaptation.
The README links a native desktop client for 3 operating systems, macOS, Windows, and Linux, plus an online experts page. It also points to a 35-lesson course focused on using the 277 experts individually and in teams. That supporting material can help newcomers because a catalog this large can become a directory people admire but rarely use. The practical win is finding a few roles that match recurring work and refining them against real examples.
The missing evaluation layer matters more than the missing build
The absence of a build is not a defect by itself because there may be nothing to compile. The absence of automated tests is more consequential. A role can be syntactically valid yet produce vague plans, unsafe actions, weak citations, or inconsistent deliverables. Our box found no test target across 366 files. The claims about independent workflows and useful outputs therefore remain documentation claims, not measured performance. Teams should create scenario checks for every role that reaches production work.
The supplied README excerpt is crowded with at least 7 sponsor blocks, discount codes, API intermediaries, courses, and companion products. Sponsorship is not evidence against the library, but it makes project instructions harder to scan. Security-conscious teams should be careful around third-party model gateways: nothing in our no-secrets run tested those services, their pricing, their uptime, or their handling of prompts and credentials. Treat commercial claims as advertisements, not as project capabilities we verified.
Current activity is stronger evidence than the June release
The repository had 20,550 stars and only 1 open issue at review time. More importantly, its last push was September 10, 2026, the day of our review. The latest tagged release was v1.2.6 on June 16, 2026. That gap does not suggest abandonment because the source is still moving. The low issue count and same-day push together look healthy, although one issue total cannot show how quickly maintainers respond or whether users report problems elsewhere.
The community signal is unusually strong for a project created March 6, 2026: the supplied community item recorded 20,497 stars, while the repository snapshot showed 20,550. Fast adoption can attract contributions and speed translations, but it can also amplify uneven material. With 277 roles and no automated behavioral suite, maintainers need editorial discipline as much as code review. Version tags identify the bundle; they do not certify that a compliance or finance persona is accurate enough for unsupervised use.
It fits beside your assistant, with your controls around it
In a real stack, treat agency-agents-zh as a source layer. Select 1 role, inspect its instructions, adapt its deliverable format, and load it into a supported tool. Keep model access, retrieval, permissions, approvals, logging, and evaluation in your existing platform. For higher-risk work, require human review and restrict tool access. The project can help define who the assistant is supposed to be, but it does not provide the runtime controls that determine what the assistant may do.
Start with 2 or 3 recurring tasks and compare the selected roles with a plain baseline prompt. Record missed requirements, unsupported claims, and intervention rates, then keep only definitions that improve actual work. If your team is English-first, the upstream repository is the cleaner default. If your work depends on Chinese platforms or local business vocabulary, this edition is more practical, provided you regard its 277 roles as editable operating drafts rather than certified experts.