The Agency

The Agency is a prompt engineering toolkit, not a full agent system. It bundles 232 structured expert role files (Markdown) covering engineering, design, marketing, legal, and more. If you already use LLM-powered coding assistants like Claude Code, Cursor, or Gemini CLI and need multi-perspective design reviews, code reviews, or documentation, this library is a low-cost experiment. However, real-world adoption evidence comes mostly from a few individual tests with low upvotes—treat it as a productivity experiment rather than a replacement for a human cross-functional team.

In a single tester’s write-up, injecting roles slashed architecture design from 1–2 days to 4–6 hours, boosted test coverage from 40% to over 80%, and cut per-feature code review issues from 15–20 down to 3–5. These numbers are from one Zhihu post (2 upvotes), so the sample is extremely small. Roles work by enforcing identity, skills, workflow, and deliverable standards, keeping the LLM’s output focused and contextually coherent. Usage is simple: clone the repo, run the install script, pick departments, and activate roles inside your assistant.