Agency Agents is a role library, not evidence that more agents win.
The useful part of a 264-role catalog is not pretending you hired an agency. It is selecting two or three narrow jobs, defining their artifacts, and making the workflow earn its extra cost against a one-agent baseline.
You leave with A source review plus two pullable workflows: a three-role code-change gate and a controlled one-agent-versus-multi-agent experiment
Evidence label: source review, not benchmark
We reviewed the Agency Agents repository at commit 6d29a9b on September 9, 2026. We inspected its installer, converter, contribution rules, workflow example, and the role files used below. We did not benchmark its roles against unmodified prompts, reproduce every supported integration, or evaluate the desktop app. Software Sausage has no affiliation with the project.
What the project actually provides
The MIT-licensed repository contains 264 source agent definitions across 18 divisions. Each Markdown file combines a specialized persona with deliverables, a process, boundaries, and success metrics. Conversion and installation scripts target Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Qwen Code, Aider, and other harnesses.
The most important installer feature is selection. It can install one agent or a division, show a dry run, and write to an explicit path. That matters more than the catalog size. The project itself warns that OpenCode currently registers only about 119 agents and recommends subsets.
At review time, GitHub reported roughly 151,000 stars and 24,000 forks. The latest commit's visible checks for installer behavior, tool and division manifests, and runbook slugs were green. Those are strong maintenance and packaging signals; they do not measure whether the agent prompts improve task results.
The repository's best rule is in CONTRIBUTING
The contribution guide requires a narrow specialization, distinct behavior, concrete deliverables, measurable success, and real testing. It also rejects near-duplicate re-skins and generated integration output in pull requests.
That standard points away from “activate the whole agency.” If two roles produce the same artifact, if nobody consumes an output, or if success cannot be checked, the role should not be in the workflow.
Where the claims outrun the evidence
- The README's “never sleep” and “always deliver” language is positioning, not a reliability result. Agent prompts still time out, misunderstand context, and produce plausible errors.
- A large roster creates selection and context costs. A prompt file does not create process isolation, least privilege, or an independent reviewer.
- The seven-agent startup example recommends passing full outputs between roles. The newer multi-agent architect instead warns about context growth and recommends summaries and structured state. The second rule is safer.
- Green repository checks prove the catalog and installers are internally consistent. They do not show that seven agents beat one capable agent on correctness, cost, or elapsed time.
Two smaller recipes
| Job | Roles | Gate |
|---|---|---|
| Ship one code change | Implementer → reviewer → security auditor | Repository checks, fresh runtime proof, human decision |
| Decide if orchestration helps | One-agent baseline ↔ no-more-than-three-role trial | Same fixture, verifier, blind rubric, and predeclared threshold |
The code-change recipe combines the project's Minimal Change Engineer, Code Reviewer, and AI-Generated Code Security Auditor. Only the implementer writes code. The other roles receive the diff and check logs, and a person makes the release decision.
The experiment recipe starts with one agent, permits at most three roles in the challenger, caps retries at two, and runs both arms against the same fixture and verifier. Multi-agent stays only if it clears a threshold chosen before the result is known.
Pull the ledgers, then try to disprove them
Both kits include an editable run ledger. Their checks prove required decision fields remain present, not that the workflow improves your job. Run the experiment on a disposable fixture and publish failed runs too.
Get the measured result
One maintained workflow at a time. Explicit opt-in, no vendor-paid ranking, unsubscribe any time.