Agent System – Orquestração Multi-agente com Obsidian
GaspardCoche/agent-system
Orquestração de agentes IA com memória persistente em Obsidian e GitHub Actions
Ver no GitHubSistema multi-agentes autônomo orquestrado via GitHub Actions e Claude Code, onde cada agente especializado (Pesquisador, Analisador, Forjador, Sentinel) possui memória persistente em Obsidian Vault e comunica via protocolo padronizado. Inclui 10 agentes com papéis específicos (lideração, análise, automação, auditoria) e knowledge graph integrado para captura de aprendizados.
Instalação
npx convez-ai GaspardCoche/agent-system1 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
Casos de uso
- Automação de workflows complexos com múltiplos agentes paralelos
- Gestão de conhecimento empresarial com knowledge graph interativo
- Auditoria, triage de dados e geração de insights automatizados
- CI/CD avançado com validação, testes e agregação de relatórios
Veredito

Arquitetura empresarial robusta para orquestração de agentes com persistência de conhecimento. Excelente para times que precisam coordenar múltiplos bots autônomos com contexto compartilhado; requer experiência em GitHub Actions, Python e design de sistemas distribuídos.
README
A tradução para o português deste README ainda está na fila.
Agent System
Systeme multi-agents autonome orchestre via GitHub Actions et Claude Code. Chaque agent a un role specialise, une memoire persistante (Obsidian Vault), et un protocole de communication standardise.
Architecture
GitHub Issues / Cron / Manual
|
Orchestrator ──── decompose la tache
|
+----+----+
| |
Researcher Analyzer (parallele)
| |
+----+----+
|
Forge ──── implemente
|
Sentinel ──── teste & valide
|
Aggregate ──── rapport + notification
Agents
| Agent | Role | Trigger |
|---|---|---|
| Orchestrator | Decompose, route, supervise | Issues, PRs, cron, manual |
| Forge | Implementation code | Dispatch |
| Sentinel | Tests & QA | Dispatch |
| Iris | Email triage & digest | Cron 7h30, manual |
| Scout | Veille web, enrichissement | Dispatch, manual |
| Aria | Lead generation, CRM | Dispatch, manual |
| Nexus | Audit Google Ads | Cron Lundi, manual |
| Sage | Auto-amelioration systeme | Cron Dimanche, manual |
| Ralph | Automatisation, webhooks | repository_dispatch |
| Lumen | Analyse donnees & insights | Dispatch |
Vault (Knowledge Base)
Le vault Obsidian dans docs/vault/ est la memoire persistante du systeme. Chaque agent :
- Lit
INDEX.md+ son fichier memoire au debut de chaque run - Ecrit ses apprentissages apres chaque run
- Le knowledge graph est automatiquement reconstruit via
vault-sync.yml
Structure du Vault
docs/vault/
INDEX.md -- Hub central, point d'entree
business/
vision.md -- Vision & mission
strategy.md -- Strategie business
roadmap.md -- Roadmap & priorites
prospects/
pipeline.md -- Pipeline leads (Scout + Aria)
campaigns/
google-ads.md -- Etat campagnes (Nexus)
operations/
daily-digest.md -- Digest email (Iris)
decisions.md -- Decisions d'architecture
runbooks.md -- Procedures operationnelles
kpis.md -- Metriques systeme
content/
social-media.md -- Strategie contenu
brand-voice.md -- Ton editorial
editorial-calendar.md -- Calendrier editorial
tech/
infrastructure.md -- Architecture technique
integrations.md -- APIs & MCP servers
security/
access-control.md -- Controle d'acces
agents/
dispatch-log.md -- Journal orchestrations
nexus-memory.md -- Memoire Nexus
iris-memory.md -- Memoire Iris
scout-memory.md -- Memoire Scout
sage-memory.md -- Memoire Sage
forge-memory.md -- Memoire Forge
sentinel-memory.md -- Memoire Sentinel
ralph-memory.md -- Memoire Ralph
lumen-memory.md -- Memoire Lumen
aria-memory.md -- Memoire Aria
templates/
vault-note.md -- Template note generique
agent-memory.md -- Template memoire agent
Ouvrir dans Obsidian
open -a Obsidian docs/vault
Obsidian affiche le graph interactif des connexions entre les notes (Cmd+Shift+G).
Dashboard
Dashboard web deploye sur GitHub Pages (https://gaspardcoche.github.io/agent-system/) avec 4 onglets :
- Runs -- Historique des runs, lancement d'agents, rapports inline
- Status -- Sante des agents, secrets configures/manquants
- Agents -- Documentation, inputs, instructions de setup
- Graph -- Knowledge graph interactif (D3.js)
Skills
Systeme de skills qui transforme les patterns MCP frequents en scripts Python autonomes :
| Skill | Description |
|---|---|
firecrawl_scrape | Scrape URL -> markdown |
github_create_issue | Creer/commenter issues |
gemini_analyze | Analyse gros fichiers via Gemini |
slack_notify | Notification Slack formatee |
Les skills sont geres par Sage qui analyse les retrospectives et propose de nouveaux skills.
Setup
Prerequis
- GitHub repo avec Actions activees
- Secrets GitHub configures (voir ci-dessous)
- GitHub Pages active sur le repo (deploy depuis
docs/)
Secrets GitHub requis
| Secret | Agent(s) | Notes |
|---|---|---|
CLAUDE_CODE_OAUTH_TOKEN | Tous | claude setup-token en local |
GEMINI_API_KEY | Lumen, Sage | Google AI Studio |
GMAIL_TOKEN_JSON | Iris | OAuth Gmail |
GMAIL_USER_EMAIL | Iris | Adresse Gmail |
FIRECRAWL_API_KEY | Scout | firecrawl.dev |
HUBSPOT_API_KEY | Aria | HubSpot private app |
FULLENRICH_API_KEY | Aria | fullenrich.com |
GOOGLE_ADS_DEVELOPER_TOKEN | Nexus | Google Ads API Center |
GOOGLE_ADS_CLIENT_ID | Nexus | Google Cloud OAuth |
GOOGLE_ADS_CLIENT_SECRET | Nexus | Google Cloud OAuth |
GOOGLE_ADS_REFRESH_TOKEN | Nexus | OAuth flow |
SLACK_WEBHOOK_URL | Notifications | Slack Incoming Webhook |
Lancer un agent
# Via GitHub CLI
gh workflow run orchestrator.yml -f task_description="Description de la tache"
# Agent specifique
gh workflow run nexus.yml -f report_type=weekly_audit -f dry_run=true
# Via le dashboard GitHub Pages
# Ouvrir https://gaspardcoche.github.io/agent-system/ -> Config (PAT) -> Run agent
Rebuild le knowledge graph
python3 .github/scripts/vault_builder.py --vault-dir docs/vault --output docs/data/graph.json
Protocole de communication inter-agents
Chaque agent lit sa tache depuis /tmp/agent_task.json et ecrit ses resultats dans /tmp/agent_result.json :
{
"agent": "<nom>",
"task_id": "<id>",
"status": "complete|failed|needs_retry|pending_approval",
"summary": "<max 150 mots>",
"artifacts": ["<filepath>"],
"next_agent": "<nom ou null>",
"retrospective": {
"what_worked": "...",
"what_failed": "...",
"mcp_patterns": ["<tool:pattern:count>"],
"improvement_suggestion": "..."
}
}
Self-Improvement
- Chaque agent remplit
retrospectivedans son resultat JSON - Sage lit toutes les retrospectives chaque dimanche
- Sage propose des ameliorations de prompts via PR
- Les erreurs repetees vont dans
memory/lessons_learned.md - Les patterns MCP frequents deviennent des skills valides
Fichiers cles
| Fichier | Role |
|---|---|
CLAUDE.md | Instructions globales pour tous les agents |
.github/workflows/_reusable-claude.yml | Base workflow commune |
.github/scripts/vault_builder.py | Genere le knowledge graph |
.github/scripts/generate_report.py | Genere les rapports |
docs/index.html | Dashboard complet |
docs/data/agents.json | Config agents pour le dashboard |
skills/registry.json | Registry des skills |
memory/lessons_learned.md | Erreurs a ne pas repeter |
Licence
Projet prive.