convez.ai
SkillAutomaçãoAvançado
daehounan

Agency Agents Fork - 163 Agentes Especialistas com Roteamento Inteligente

daehounan/agency-agents-fork

163 agentes especializados + roteamento automático para tarefas complexas em múltiplos domínios

Ver no GitHub
BomPluginScore · de 0 a 100

Coleção de 163 agentes Claude especializados em 15 domínios (desenvolvimento de games, negócios coreano/japonês, mídia paga, XR/spatial), com 24 skills de roteamento que distribuem automaticamente tarefas ao especialista correto. Inclui navegadores de negócios culturais, playbooks de estratégia e arbitrador central para desambiguação entre ~500 skills do ecossistema Claude Code.

Instalação

$npx convez-ai daehounan/agency-agents-fork

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
1
Acompanhando
0
Saúde
85%
Último commit
há 5 meses
Versão
v1.3.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Negociações com contexto cultural (Coreia/Japão) com interpretação de sinais sociais sutis
  • Desenvolvimento de games em múltiplas engines (Unity, Unreal, Godot, Roblox, Blender)
  • Otimização de campanhas de mídia paga (Google, Meta, LinkedIn, TikTok) com especialistas em routing
  • Desenvolvimento imersivo (XR, visionOS, WebXR) com agentes dedicados
  • Automação de workflows complexos que exigem múltiplos especialistas em sequência

Veredito

Arquitetura robusta de agentes multi-especialistas com foco em roteamento inteligente e contexto cultural. Excelente para workflows empresariais complexos, mas requer compreensão profunda de PowerShell 7+ e gestão de ~8.400 tokens de contexto. Fork ativo com subconjuntos otimizados disponíveis.

README

A tradução para o português deste README ainda está na fila.

Agency Agents (Fork)

163 Claude Code specialist agents + 24 routing skills, one-line install. Korean / Japanese business culture navigators · game-dev across Unity / Unreal / Godot / Roblox / Blender · XR / spatial · paid-media · skill-routing-arbitrator for disambiguating across the ~500-skill Claude Code ecosystem.

Lint Skills Build Plugin License: MIT Agents Skills Playbooks Security Policy Cross-platform Tooling

Install in one line

claude --plugin-url https://github.com/daehounan/agency-agents-fork/releases/download/v1.3.0/agency-agents-fork-v1.3.0-full.zip

Or grab a smaller scoped subset from the v1.3.0 release: engineering-finance (32 agents) · marketing-paid-media-sales (30) · game-development (20).

Just want the Korean + Japanese Business Navigators? → standalone plugin at daehounan/k-jp-business-navigators (2 agents, ~374 routing tokens vs full bundle's ~8,400).

What this actually does — Korean / Japanese Business Navigators

Q: 거래처 부장님이 첫 미팅 직후 다음 주 회식 초대했어요. 거절하면 실례인가요?

korean-business skill → Korean Business Navigator: 첫 미팅 직후 회식 초대 = strong-interest 신호. 거절하면 deal cools ~30%. KakaoTalk reply: "부장님, 초대해주셔서 감사합니다. 다음 주 [요일] 저녁 비워두겠습니다. 잘 부탁드립니다." → 2차 갈 의향 + 페이스 맞춰 따라가는 protocol 암시. 다음날 차사발 준비, 결재라인이 내려오는 시점 monitor.

Q: 取引先から「検討させていただきます」と返事が来ました。前向きですか?

japanese-business skill → Japanese Business Navigator: 80% 정중한 거절. 「前向きに検討」/「持ち帰って検討」과는 다른 valence. 14일간 후속 연락 없으면 dead — re-engage 시 graceful 言い訳 framing 필요 (e.g. 追加資料 보내는 명분). Golden Week / Obon / 年末年始 직전이면 의도와 무관하게 stall.

Same routing pattern applies to game-dev (Unity/Unreal/Godot/Roblox/Blender), paid-media (Google/Meta/LinkedIn/TikTok), XR (visionOS/WebXR), and the 14 other divisions — ask in plain language, the right specialist auto-fires.


A curated fork of msitarzewski/agency-agents — 163 specialist agent personas across 15 divisions + 24 routing skills (consolidated, no duplicates with existing ecc/Anthropic skills) + 16 NEXUS strategy playbooks + 2 hooks + ecosystem-wide duplicates audit.

Three install paths supported:

PathWhenHow
Pre-built plugin from release (recommended)You just want the plugin, no clone, no build step.claude --plugin-url https://github.com/daehounan/agency-agents-fork/releases/download/v1.3.0/agency-agents-fork-v1.3.0-full.zip
User-level config bundleYou want agents in ~/.claude/agents/ and skills in ~/.claude/skills/ (no namespacing). Windows.pwsh scripts/install.ps1 -WithSkills [-WithHooks]
Plugin built from sourceYou want to customize divisions or hack on the build.pwsh scripts/build-plugin.ps1 → claude --plugin-dir ./dist/plugin

The plugin build flattens the division-organized source into a spec-conformant agents/ directory plus hooks/hooks.json and a clean .claude-plugin/plugin.json. The division dirs in the source repo are kept for browsing. See Plugin build below.

Pre-built release assets (v1.3.0)

VariantAgentsURL suffix (under …/releases/download/v1.3.0/)
Full bundle163agency-agents-fork-v1.3.0-full.zip
Engineering + finance32agency-agents-fork-v1.3.0-engineering-finance.zip
Marketing + paid-media + sales30agency-agents-fork-v1.3.0-marketing-paid-media-sales.zip
Game development20agency-agents-fork-v1.3.0-game-development.zip

For other subsets, build locally with scripts/build-plugin.ps1 -Divisions <list>.

What changed vs upstream

ChangeDetail
China-market agents removed22 files dropped (see excluded list)
Japanese Business Navigator addedNew 216-line agent: 稟議 / 根回し / KY / 飲み会 / 報連相
24 routing skills addedAuto-invoke skills that delegate to the right specialist (see Skills). Consolidated to avoid duplicates with ecc:*, engineering:*, design:*, operations:* etc. Includes skill-routing-arbitrator for ecosystem-wide disambiguation.
Ecosystem duplicates auditdocs/skill-ecosystem-duplicates.md maps ~500 skills into preference clusters (versioned snapshots, legacy shims, cross-namespace duplicates, hooks)
UserPromptSubmit hookOptional regex-based agent-suggestion hook (hooks/suggest-agents.ps1)
Windows toolingPowerShell install.ps1 / convert.ps1 / migrate-ecc.ps1 replace bash scripts
Overlap mappingSee docs/ecc-overlap.md — which fork agents duplicate existing ecc plugins
Original count200 markdown files → 178 (China removed) → 179 (Japanese Business Navigator added). Of those 179, 163 are agent personas (with YAML frontmatter) and 16 are NEXUS strategy playbooks (documentation, no frontmatter).

Upstream license (MIT) preserved. See LICENSE.

Division roster (post-exclusion)

DivisionAgents
specialized37
engineering27
game-development20
marketing15
design8
sales8
testing8
paid-media7
project-management6
spatial-computing6
support6
academic5
finance5
product5
Agent total163

Strategy playbooks (non-agent)

The strategy/ directory contains 16 NEXUS engagement documents (not agents — no YAML frontmatter, not loadable as personas). They are referenced by the strategy-nexus routing skill:

  • EXECUTIVE-BRIEF.md, QUICKSTART.md, nexus-strategy.md
  • playbooks/phase-{0..6}-*.md — 7 phase playbooks (discovery → operate)
  • runbooks/scenario-{enterprise-feature, incident-response, marketing-campaign, startup-mvp}.md
  • coordination/{agent-activation-prompts, handoff-templates}.md

Quick start (Windows / PowerShell)

# Install 163 agents + 16 playbook docs (179 .md files total) to ~/.claude/agents
.\scripts\install.ps1

# Install just one division (legacy single-division flag)
.\scripts\install.ps1 -Division engineering

# Install a scoped subset (multi-division — UX parity with build-plugin.ps1 -Divisions)
.\scripts\install.ps1 -Divisions engineering,finance
.\scripts\install.ps1 -Divisions specialized   # Korean + Japanese Business Navigators
.\scripts\install.ps1 -Divisions "marketing;paid-media;sales"

# Dry run (see what would happen, no copies)
.\scripts\install.ps1 -DryRun

# Don't overwrite existing files
.\scripts\install.ps1 -SkipExisting

# Custom target
.\scripts\install.ps1 -Target "C:\custom\path"

# Also register the suggest-agents + skill-telemetry hooks in ~/.claude/settings.json
.\scripts\install.ps1 -WithHooks

Convert to other tools

.\scripts\convert.ps1 -Tool cursor      # -> integrations/cursor/*.mdc
.\scripts\convert.ps1 -Tool aider       # -> integrations/aider/CONVENTIONS.md
.\scripts\convert.ps1 -Tool windsurf    # -> integrations/windsurf/.windsurfrules
.\scripts\convert.ps1 -Tool opencode    # -> integrations/opencode/*.md
.\scripts\convert.ps1 -Tool all         # all four

Excluded agents

22 files removed (China-market focus):

Engineering

  • engineering-feishu-integration-developer.md
  • engineering-wechat-mini-program-developer.md

Marketing (15)

  • marketing-baidu-seo-specialist.md
  • marketing-bilibili-content-strategist.md
  • marketing-china-ecommerce-operator.md
  • marketing-china-market-localization-strategist.md
  • marketing-cross-border-ecommerce.md
  • marketing-douyin-strategist.md
  • marketing-kuaishou-strategist.md
  • marketing-livestream-commerce-coach.md
  • marketing-podcast-strategist.md (Chinese podcast market)
  • marketing-private-domain-operator.md
  • marketing-short-video-editing-coach.md
  • marketing-wechat-official-account.md
  • marketing-weibo-strategist.md
  • marketing-xiaohongshu-specialist.md
  • marketing-zhihu-strategist.md

Specialized (5)

  • government-digital-presales-consultant.md (China ToG)
  • healthcare-marketing-compliance.md (China healthcare advertising)
  • supply-chain-strategist.md (China manufacturing ecosystem)
  • study-abroad-advisor.md (Chinese students)
  • recruitment-specialist.md (China hiring platforms)

If you want any of these back, copy them in from a fresh clone of the upstream agency-agents repository — this fork removed them.

Agent file format

Each agent is a Markdown file with YAML frontmatter:

---
name: Frontend Developer
description: Expert frontend developer specializing in...
color: cyan
emoji: 🖥️
vibe: One-line vibe statement.
---

# <Agent Name> Agent Personality

## 🧠 Your Identity & Memory
...

## 🎯 Your Core Mission
...

## 🚨 Critical Rules You Must Follow
...

## 📋 Your Technical Deliverables
... (with code examples)

## 🔄 Your Workflow Process
...

## 💭 Your Communication Style
...

Skills

24 routing skills in skills/ auto-invoke based on description matching, delegating to the appropriate specialist agent(s). The original 26 were consolidated — duplicates with existing skills (engineering:*, design:*, ecc:autonomous-agent-harness, ecc:mcp-server-patterns, etc.) were removed (3 dropped, 3 slimmed) so each routing skill has genuine net-new value.

Consolidation history

Removed (3)WhyUse instead
general-engineering-routing100% overlap with engineering:* (10 skills) + ecc agentsecc:architect, ecc:code-reviewer, engineering:architecture, engineering:debug
ai-infra-routingOverlap with ecc:mcp-server-patterns, ecc:autonomous-agent-harness, anthropic-skills:mcp-builder, ruflo-swarm:*Those skills directly
design-routingOverlap with design:* (7 skills)design:design-critique, design:design-system, etc.
Slimmed (3)Original scopeNow covers only
compliance-audit-routingSOC 2/ISO/HIPAA + smart contract + n8nSmart contract + n8n only (SOC 2 etc. → operations:compliance-tracking, anthropic-skills:security-audit-engine)
knowledge-content-routingZettelkasten + doc generation + tech writerZettelkasten + corp training only (formats → anthropic-skills:pptx/docx/pdf/xlsx)
product-discovery-routing5 product agentsBehavioral nudge + trend researcher persona only (sprint/PRD/roadmap → product-management:*)

Install skills

.\scripts\install.ps1 -WithSkills              # install agents + skills
.\scripts\install.ps1 -SkillsOnly              # install only skills (~/.claude/skills/)
.\scripts\install.ps1 -WithSkills -WithHooks   # agents + skills + hooks (full install)

Skill roster

SkillRoutes toDomain
korean-businessKorean Business Navigator품의 / nunchi / 회식 / KakaoTalk
japanese-businessJapanese Business Navigator稟議 / 根回し / KY / 飲み会 / 報連相
localization-cultural3 agentsFrench ESN / cultural intel / ES↔EN translation
xr-spatial-routing6 agentsvisionOS / WebXR / Vision Pro / Metal
game-development-routing19 agentsUnity / Unreal / Godot / Roblox / Blender
niche-engineering-routing9 agentsembedded / voice AI / email intel / SIEM / CMS / AI remediation
general-engineering-routing18 agentsfrontend / backend / mobile / SRE / DevOps
paid-media-routing7 agentsGoogle Ads / Meta / programmatic / tracking
social-platform-routing11 agentsTwitter / TikTok / IG / Reddit / LinkedIn / YouTube
sales-methodology-routing9 agentsMEDDPICC / SPIN / discovery / proposal / coach
legal-firm-ops3 agentsintake / billing / document review
real-estate-mortgage2 agentsbuyer-seller / loan officer
customer-service-vertical4 agentsgeneral / healthcare / hospitality / retail
finance-analysis5 agentsbookkeeper / FP&A / DCF / tax
worldbuilding-rigor5 agentsanthropologist / geographer / historian / narratologist / psychologist
qa-testing-routing8 agentsevidence / reality / perf / API / a11y
compliance-audit-routing2 agents (slim)smart contract + n8n governance only
knowledge-content-routing2 agents (slim)Zettelkasten + corporate training only
data-ops-routing3 agentssales extraction / consolidation / report dispatch
product-discovery-routing2 agents (slim)behavioral nudge + trend researcher only
project-management-routing6 agentsproducer / shepherd / Jira / experiment
support-routing12 agentssupport / analytics / infra / civil eng / accounts payable
strategy-nexusNEXUS playbook7 phases + 4 scenarios + handoff protocols
agency-rosterDiscovery utilityBrowse the full 163-agent collection
skill-routing-arbitratorMeta-skillResolves which skill to use when multiple match (covers all ecosystem clusters)

Hook

hooks/suggest-agents.ps1 is a UserPromptSubmit hook (manual activation, not auto-enabled). Silent on no-match, injects one-line agent hint on high-confidence regex pattern matches. See hooks/README.md for activation JSON.

Overlap with existing ecc / Anthropic skills

Many fork agents duplicate capabilities already provided by your ecc plugin and Anthropic skills. Skills documented here cross-link to stronger alternatives (e.g., general-engineering-routing defers to ecc:architect and ecc:code-reviewer for those use cases). See docs/ecc-overlap.md for the side-by-side mapping.

Plugin build

scripts/build-plugin.ps1 produces a Claude Code plugin-loader-conformant build at dist/plugin/ from the division-organized source.

.\scripts\build-plugin.ps1                                # all 163 agents
.\scripts\build-plugin.ps1 -Divisions finance,engineering # scoped subset (32 agents)
.\scripts\build-plugin.ps1 -Divisions game-development    # single division (20 agents)
.\scripts\build-plugin.ps1 -Verbose_                      # show each file as it's copied
.\scripts\build-plugin.ps1 -Zip                           # also emit dist/agency-agents-fork-plugin.zip

About token cost

The full 163-agent bundle's agent description fields sum to ~8k tokens by static estimation, but Claude Code's runtime warning fires at ~22k cumulative — it counts more than just description fields. That's above Claude's recommended 15k budget for plugin-supplied agent metadata, so every turn pays a real overhead.

If you only need a slice of the collection, build a scoped plugin:

Use caseCommandAgentsPlugin name
Game dev only-Divisions game-development20agency-agents-fork-game-development
Engineering + finance-Divisions engineering,finance32agency-agents-fork-engineering-finance
Marketing + paid-media + sales-Divisions marketing,paid-media,sales30agency-agents-fork-marketing-paid-media-sales
Korean/Japanese business prep-Divisions specialized37agency-agents-fork-specialized
Full bundle (default)(no flag)163agency-agents-fork

Valid divisions: academic, design, engineering, finance, game-development, marketing, paid-media, product, project-management, sales, spatial-computing, specialized, support, testing.

Comma, semicolon, and whitespace all work as separators: -Divisions finance,engineering or -Divisions "finance;engineering;academic" are both fine.

All 24 routing skills are always included regardless of scope — they're small (~14k chars total) and several route to skills outside the local bundle (e.g. ecc:*, anthropic-skills:*), so they remain useful even in a slim build.

Resulting structure:

dist/plugin/
├── .claude-plugin/plugin.json   # manifest only (no file lists — spec-conformant)
├── agents/                      # 163 flat .md files (flattened from division dirs)
├── skills/                      # 24 skill directories with SKILL.md
├── hooks/
│   ├── hooks.json               # declares UserPromptSubmit + PreToolUse:Skill
│   ├── suggest-agents.ps1
│   └── log-skill-fired.ps1
└── README.md

Test the build locally — run from the repo root (cd first), or use absolute paths:

# Option A: cd to repo first (recommended)
cd C:\Users\andae\Projects\agency-agents-fork
.\scripts\build-plugin.ps1
claude --plugin-dir .\dist\plugin

# Option B: absolute paths from anywhere (e.g. another terminal session)
pwsh -NoProfile -File C:\Users\andae\Projects\agency-agents-fork\scripts\build-plugin.ps1
claude --plugin-dir C:\Users\andae\Projects\agency-agents-fork\dist\plugin

If you run .\scripts\build-plugin.ps1 from C:\Users\andae (or any directory that isn't the repo root), PowerShell can't find the script — that's not a build failure, it's the wrong working directory. The build script prints the absolute --plugin-dir path on success so you can copy-paste it regardless of cwd.

Or zip the directory and use --plugin-url:

.\scripts\build-plugin.ps1 -Zip
# emits dist/agency-agents-fork-plugin.zip plus a ready-to-paste --plugin-url command

The build is verified on every push by build-plugin.yml — asserts 163 agents, 24 skills, both JSON manifests parse, then uploads the artifact for 30-day retention.

dist/ is gitignored. Source-of-truth is the division dirs; the build is reproducible from them.

Security

See SECURITY.md for the vulnerability reporting policy, scope (PowerShell scripts, hooks, agent personas), and the hardening already in place (settings.json backup, silent-on-error hooks, no outbound network).

Verification scripts

ScriptWhat it doesRun on CI
scripts/lint-skills.ps1YAML frontmatter + description + body validation across all 24 skills✅
scripts/audit-agent-refs.ps1Cross-references every backticked slug in skills/*/SKILL.md against the actual 163 agent files (catches typos, stale refs, renamed agents)✅
scripts/build-plugin.ps1Builds a Claude Code plugin-loader-conformant package at dist/plugin/ (flat agents/ + hooks/hooks.json + clean manifest)✅
scripts/show-skill-stats.ps1Reads the local telemetry log and reports fire counts per skill—
scripts/install-hooks-in-settings.ps1Idempotent registration of UserPromptSubmit + PreToolUse(Skill) hooks in ~/.claude/settings.json with backup/rollback—

External skill names referenced from routing tables without a namespace prefix (e.g. built-in /review, internal-comms) are whitelisted in scripts/external-skill-refs.txt so the audit treats them as resolved.

License

MIT (preserved from upstream).

Ver o original no GitHub

Bom para

Relacionados