The Agency of 90 Agents Runs on 82% Shell Script
agency-agents is a collection of specialized AI agent personas, and it's popular — 141,235 stars, still getting commits as of three days ago. But the engineering shape underneath is thin: no releases to pin to, no visible test suite, and a language mix (82% Shell) that says more about how this thing is glued together than the README does. Star count and code maturity are two different questions, and here the gap between them is wide.
| Last push | 3 days ago |
| Stars / Forks | 141,235 / 23,041 |
| Created | 300 days ago |
| License | MIT (permissive, non-copyleft) |
| Language mix | Shell 82% · Python 17% · PowerShell 1% |
| Contributors | 90 (top contributor: 43% of commits) |
| Recent commits | 100 |
| Releases tagged | 0 |
| Open issues | 110 |
| File count | 343 |
| Test directory / CI / Dockerfile | Absent / Present / Absent |
The build materials
Open the file tree on a repo that markets itself as "a complete AI agency" and you'd expect a Python or TypeScript core — something with an orchestration layer, a framework underneath the agent personas. That's not what's here. Shell accounts for 82% of the codebase, Python for 17%, and PowerShell for a token 1%. Whatever coordination is happening between these agents, it's happening mostly through shell scripts calling out to something else, not through an application framework.
That's not automatically a defect — Shell is the honest choice for a repo that's really a collection of prompts, configs, and glue around an existing CLI. But it changes what "reading the code" means before you adopt it. You're not stepping into a Python package with type hints and an import graph. You're stepping into scripts, and scripts are where platform assumptions (paths, shells, environment variables) hide until they break on someone else's machine. The 1% PowerShell suggests someone attempted Windows compatibility; it doesn't tell you how complete that attempt is.
A codebase that's four-fifths Shell isn't a red flag by itself, but it means your due diligence has to be reading actual scripts line by line — there's no type system or framework contract doing that work for you.
No version to point at
Three hundred days old, still shipping — 100 commits show up in the recent window, last push three days back — and zero tagged releases. Not one. If you want to depend on this repo in anything you'd call production, there's no v1.0 or v0.4.2 to reference. Your only option is pinning to a commit SHA, which works but means every update is a manual diff against a moving target, not a changelog you can read.
Ninety people have contributed commits, with the top contributor responsible for 43% of them. That's a commit-authorship split, nothing more — it doesn't tell you who reviews pull requests, who else understands the trickier scripts, or what happens if the top contributor steps away. Treat it as one data point, not a governance chart.
The parts nobody's checked
There's a CI workflow configured — that much is confirmed. What it actually verifies, and whether it's currently green, wasn't checked here; a workflow file proves a pipeline exists, not that it passes. There's no test directory anywhere in the 343 files. That doesn't mean nothing is tested — some projects verify by other means — but there's no visible test suite for a reader to inspect, and 110 open issues sitting against a young, fast-moving repo is worth watching rather than ignoring.
No test directory and no tagged release together mean you inherit the verification work. Before this touches anything you'd call production, you're the QA department.
Questions people ask
Is agency-agents actively maintained?
Yes, by activity — the last push was 3 days ago (checked August 10, 2026) and 100 commits show up in the recent window, against a repo that's 300 days old.
Can I pin to a stable version?
Not with a release tag — the repository has zero tagged releases. The only pin point available is a specific commit.
Does it have tests?
No test directory was found in the 343 tracked files. A CI workflow is configured, but whether it currently passes wasn't checked.
What is it actually written in?
Mostly Shell (82%), with Python (17%) and a small amount of PowerShell (1%) — not a Python-first agent framework, despite the framing.
The verdict
Fine for what the popularity suggests it's actually used for: browsing agent personas, borrowing prompts, running it locally out of curiosity. The MIT license means you can do that and reuse what you find, so long as you keep the license notice — it says nothing about anything this repo might pull in or depend on. Wrong for anything wired into a production pipeline without your own review pass first: no release to pin, no visible tests, and a codebase that's mostly Shell means you're reading the scripts yourself before you trust them. Pin a commit, read the parts you depend on, and don't assume the star count already did that work for you.
Read the full article
Comments
Post a Comment