Michael Sitarzewski’s The Agency — an MIT-licensed bundle of 147 specialized AI agent personalities organized across 12 operational divisions (Engineering 30+, Design 7, Paid Media 8, Sales 9, Marketing 27+, Product 5, Project Management 5, Testing 8, Support 6, Spatial Computing 5, Game Development 25+, Academic 5, Specialized 40+, Finance 5). Ships with a multi-tool installer that targets 11+ AI coding assistants: Claude Code, Cursor, GitHub Copilot, Antigravity (Gemini), Gemini CLI, OpenCode, OpenClaw, Aider, Windsurf, Qwen Code, Kimi Code. 95,079 stars / 15,683 forks at fetch (2026-05-08), making it one of the largest community AI-agent bundles in existence — created October 13 2025, last push April 12 2026, ~6 months old. Author Michael Sitarzewski is a real founder (Techstars alum, 30+ years building, 22 public repos, 3.4k GitHub followers). Closest neighbor in this wiki: agents (Claude Code marketplace, 34.4k★) — same idea (one repo, hundreds of curated agents), different shape (wshobson is a marketplace plugin with /plugin marketplace add; agency-agents is a markdown-agent collection with shell-script multi-tool installation).
Key Takeaways
147 agents organized into 12 divisions — distinct personalities, processes, and measurable success criteria. Not just one-line prompts: each agent ships with workflows, deliverables, and “Identity & Memory + Core Mission + Critical Rules + Technical Deliverables + Workflow Process + Success Metrics” structure (per the contributing template).
Multi-tool installer is the load-bearing differentiator.
./scripts/install.sh --tool claude-code for Claude Code (recommended)
./scripts/install.sh --tool cursor for Cursor
./scripts/install.sh --tool aider for Aider
./scripts/install.sh --tool opencode, --tool copilot, --tool windsurf, etc. for the rest
./scripts/install.sh (no flag) auto-detects installed tools and installs interactively
The convert.sh script generates integration files; install.sh does the placement.
Notable agent examples (not exhaustive):
Frontend Developer — React/Vue/Angular, UI implementation, Core Web Vitals optimization
Reddit Community Builder — “you’re becoming a valued community member who happens to represent a brand” — explicit value-driven engagement framing
Whimsy Injector — personality and delight; “every playful element must serve a functional or emotional purpose”
Reality Checker — evidence-based certification and quality gates; production-readiness verification
Evidence Collector — screenshot-based QA with visual proof, defaults to finding 3-5 issues per review
Region-specific agents are unusual. Marketing division includes dedicated agents for Xiaohongshu (RED/小红书), WeChat, Zhihu, Baidu, Bilibili, Kuaishou, Douyin, Weibo, Livestream Commerce, China Market Localization. Engineering division includes WeChat Mini Programs and Feishu Integration. Game Development division includes Unity, Unreal Engine, Godot, Blender, Roblox Studio specialists. Niche-by-platform/engine specialization is the differentiator.
Author identity is verified, not anonymous. Michael Sitarzewski (msitarzewski) is a Techstars alum, multiple companies, 22 public repos on a 12-year-old GitHub account (created 2012), 3,367 followers, blog at friendmichael.com. No supply-chain warning signal — fits the Skills at Depth onboarding module’s “verify the author” rule cleanly.
MIT licensed. Free for commercial and personal use; community translations to Simplified Chinese (zh-CN) maintained by @jnMetaCode and @dsclca12.
Star velocity is the strong adoption signal. ~95k stars in ~6 months from a non-celebrity author. 15.7k forks suggests genuine usage. Contrast with agents (~34k★ at ingest 2026-04, also a multi-agent bundle) — different shape, different distribution model, comparable adoption velocity.
Where this fits in the wiki
Sister bundle to agents — both are massive curated agent bundles. wshobson has 184 agents + 78 plugins + 150 skills + 98 commands across 25 domain categories, distributed via Claude Code marketplace plugin. agency-agents has 147 agents across 12 divisions, distributed via shell script across 11+ tools. Operators choosing between them should compare by: (a) which tools they use (Claude Code only → wshobson; multi-tool → agency-agents), (b) division depth that matches their work, (c) marketplace vs filesystem install preference.
Adjacent to skills (curated-by-Anthropic) and Corey Haines marketing-skills (single-domain). agency-agents is the broadest community-curated answer.
Region-specific marketing agents are useful for AI Marketing users working with cross-border or East-Asia-market campaigns. The Xiaohongshu / WeChat / Bilibili / Douyin / Kuaishou agents fill a gap none of the other bundles cover.
Game-development division is unique among curated bundles. Engine-specific agents (Unity / Unreal / Godot / Blender / Roblox) aren’t in any other bundle in the wiki. For non-coding-product gamedev workflows, this is the only curated path.
Composes with oh-my-claudecode (OMC) — OMC is the orchestration layer; agency-agents provides the personalities OMC’s /team mode can recruit from.
Pairs with Everything Claude Code (ECC) as the highest-volume Claude-Code-adjacent agent collection on GitHub. ECC ships 48 agents focused on agent-harness performance; agency-agents ships 147 agents focused on domain personalities. Different shapes, complementary stacks.
Implementation
Tool/Service: The Agency (msitarzewski/agency-agents)
Setup — Claude Code (recommended):
git clone https://github.com/msitarzewski/agency-agentscd agency-agents./scripts/install.sh --tool claude-code# Or to grab just one division:cp engineering/*.md ~/.claude/agents/
Activate in a session: Hey Claude, activate Frontend Developer mode and help me build a React component.
Cost: Free (MIT). Token costs go to the model provider you’re invoking through your AI tool.
Integration notes:
Each agent is a single markdown file with structured frontmatter + body. Forkable by design.
The convert script handles tool-specific format conversion (different tools want different agent file shapes).
Adding new agents: PR with the structured template (Identity & Memory + Core Mission + Critical Rules + Technical Deliverables + Workflow Process + Success Metrics).
Pick a subset, not the whole bundle. 147 agents in ~/.claude/agents/ will be hard to navigate. Pull just the divisions matching your team’s work.
Open Questions
Subagent integration vs personality activation. Agents-as-files-in-~/.claude/agents/ invokes Claude Code’s Subagents surface, but the README also describes “Hey Claude, activate Frontend Developer mode” as a personality-prefix pattern (which is closer to Skills-style activation). Which is the canonical invocation? Probably both, but worth verifying for the agency-agents-vs-Subagents-vs-Skills distinction.
Quality variance across 147 agents. With 12 divisions and 40+ specialized agents, the curation bar inevitably varies. The README publishes 5 use-case templates and ~10 named agent examples — those are the proven combos. Beyond those, treat each as untested.
Region-specific agent freshness. Chinese-platform agents (Xiaohongshu, Bilibili, Douyin) require platform-specific knowledge that goes stale fast. Last push 2026-04-12; what’s the maintenance cadence per agent? Worth checking before deploying for a real Chinese-market campaign.
Game-development engine specialists. How deeply versioned are these? Unity 6 / Unreal 5.4 / Godot 4.4 each have material API and toolchain differences. The README doesn’t list engine versions per agent.
Cross-tool format conversion fidelity.convert.sh translates agents to 11+ tool formats. How lossy is the conversion? Does the Claude-Code-native version drive different behavior than the Cursor-converted version of the same agent?
Subagent vs agency-agents trade-off.Subagents are Claude Code’s native primitive with scoped permissions and isolated workers. agency-agents personalities extend the same primitive with predefined identity / process / success-criteria. Worth a side-by-side article.
Try It
Install one division. Don’t grab the whole bundle. Pick the division matching the work you’re doing this week.
Run the documented “Startup MVP” flow. Frontend Developer + Backend Architect + Growth Hacker + Reality Checker + Rapid Prototyper. Compare against your normal solo Claude Code session for the same task.
For non-coding-product workflows, try the marketing division. Content Creator + Twitter Engager + Reddit Community Builder is a five-minute test against a real campaign brief.
For East-Asia-market work, try the region-specific agents. Xiaohongshu Strategist + China Market Localization is unique to this bundle.
For game development, try one of the engine specialists (Unity / Unreal / Godot / Blender / Roblox). The Game Designer + Level Designer + Technical Artist three-pack is the documented starting cluster.
Compare against agents if you also use that. Same task, both bundles, see which division’s interpretation feels right for your codebase / brand / team.
For multi-tool teams, the --tool flag matrix is the differentiator. If half your team uses Cursor and half uses Claude Code, agency-agents lets you keep one source of truth with both deployments.