Ein Agent Harness ist nicht Claude Code selbst. Es ist die Schicht drumherum, die Multi-Agent Workflows orchestriert, State Management handhabt, und Production-Ready macht.
Was ist ein Agent Harness?
Claude Code allein kann:
- Eine Task abarbeiten
- Eine Datei bearbeiten
- Ein Shell-Kommando ausführen
Ein Harness kann:
- 5+ Agents gleichzeitig coordinieren
- State zwischen Agents synchronisieren
- Fehler automatisch beheben
- Git Branches isolieren
- Test-Suites ausführen
- Monitoring & Logging
- Retry Logic
Die Top 5 Harnesses
1. Everything Claude Code (affaan-m)
GitHub: https://github.com/affaan-m/everything-claude-code
Stats: 84.000+ Stars. Die Production-Battle-Tested Lösung.
Architektur:
+-------------------+
| Multi-Agent Swarm | (bis 50 Agents parallel)
+-------------------+
↓
+-------------------+
| Memory Manager | (LLM Context History)
+-------------------+
↓
+-------------------+
| Skill Registry | (67 vordefinierte Skills)
+-------------------+
↓
+-------------------+
| Safety Hooks | (Guardrails, Sanitization)
+-------------------+
↓
+-------------------+
| Git Integration | (Auto Worktrees, Commits)
+-------------------+
Features:
- Multi-Agent Orchestration (Analyst → Builder → Reviewer)
- Memory System (Context Window Manager)
- 67 Skills vorinstalliert
- Safety Hooks (Output Sanitizer, Credential Vault)
- Instincts (Agent Personality Config)
- Auto-Git Integration
Installation:
git clone https://github.com/affaan-m/everything-claude-code.git
cd everything-claude-code
npm install
node index.js --config config.prod.json
Wann nutzen:
- Große Teams (5+ Developer)
- Production Deployment
- Komplexe Workflows
- Enterprise Security Requirements
Lernkurve: 2-3 Tage (robust, aber komplex)
2. Claude Agentic Framework (dralgorhythm)
GitHub: https://github.com/dralgorhythm/claude-agentic-framework
Stats: 8.200 Stars. Das MVP-Template.
Architektur:
+---------------------------+
| Agent Config (YAML) |
+---------------------------+
↓
+---------------------------+
| Pre-configured Swarms | (3 vordefiniert)
+---------------------------+
↓
+---------------------------+
| 67 Skills Library |
+---------------------------+
↓
+---------------------------+
| Minimal Orchestrator |
+---------------------------+
Features:
- Drop-in Template (Copy & Go)
- 3 vorkonfigurierte Agent-Teams
- 67 klassifizierte Skills
- Safety Hooks inklusive
- Docker-ready
Installation:
git clone https://github.com/dralgorhythm/claude-agentic-framework.git
cd claude-agentic-framework
npm install && npm run start
Wann nutzen:
- MVP Phase
- Schneller Go-to-Market
- 1-3 Agents
- Begrenzte Infrastructure
Lernkurve: 1 Tag (sehr accessible)
3. Ruflo (ruvnet)
GitHub: https://github.com/ruvnet/ruflo
Stats: 6.100 Stars. Enterprise Platform.
Besonderheit: Verteilte Agent Swarms über mehrere Nodes.
Architektur:
+-------------------------------------+
| Agent Swarm Manager |
| - up to 100+ agents distributed |
+-------------------------------------+
↓
+-------------------------------------+
| RAG (Retrieval-Augmented Gen) |
| - Knowledge Base Integration |
+-------------------------------------+
↓
+-------------------------------------+
| Real-Time Dashboard |
| - Agent Status, Resource Monitor |
+-------------------------------------+
↓
+-------------------------------------+
| Claude Code as First-Class |
+-------------------------------------+
Features:
- Verteilte Multi-Agent Orchestration
- RAG Integration (Knowledge Bases)
- Real-Time Monitoring Dashboard
- Claude Code native (kein Wrapper)
- Horizontal Scaling
Installation:
git clone https://github.com/ruvnet/ruflo.git
cd ruflo
docker-compose up -d
# Dashboard: http://localhost:3000
Wann nutzen:
- Große Scale (50+ Agents)
- Knowledge-Base Integration
- Real-Time Monitoring notwendig
- Enterprise Deployment
Lernkurve: 3-4 Tage (anspruchsvoll)
4. Composio Agent Orchestrator
GitHub: https://github.com/ComposioHQ/agent-orchestrator
Stats: 5.900 Stars. Developer-Friendly.
Besonderheit: Parallele Agents mit automatischen PRs.
Architektur:
+------------------------+
| Parallel Agents | (3+ gleichzeitig)
+------------------------+
↓
+------------------------+
| Git Worktree Manager | (pro Agent separate Branch)
+------------------------+
↓
+------------------------+
| Conflict Resolver | (automatisch)
+------------------------+
↓
+------------------------+
| PR Generator | (auto-create PRs)
+------------------------+
↓
+------------------------+
| Test Integration | (run tests before merge)
+------------------------+
Features:
- Parallele Agent-Ausführung
- Per-Agent Git Worktrees
- Automatische PR-Generierung
- Conflict Resolution
- Test Integration
Installation:
pip install composio-agent-orchestrator
composio init
composio add claude-code
composio run --agents 3 --parallel
Wann nutzen:
- Parallel Development notwendig
- Git-native Teams
- Auto-PR Workflows
- Entwickler-orientiert
Lernkurve: 1-2 Tage
5. Your Claude Engineer (coleam00)
GitHub: https://github.com/coleam00/your-claude-engineer
Stats: 4.200 Stars. Integrations-Focus.
Besonderheit: Slack, GitHub, Linear Integration.
Architektur:
+--------------------+
| Slack Interface |
+--------------------+
↓
+--------------------+
| Claude Code Agent |
+--------------------+
↓
+--------------------+
| GitHub Integration |
+--------------------+
↓
+--------------------+
| Linear Integration |
+--------------------+
Features:
- Slack Chat Interface
- GitHub Issue/PR Integration
- Linear Integration (Tasks)
- Natural Language Commands
Installation:
git clone https://github.com/coleam00/your-claude-engineer.git
cd your-claude-engineer
# Config: SLACK_TOKEN, GITHUB_TOKEN, LINEAR_TOKEN
npm run start
Wann nutzen:
- Slack-first Teams
- GitHub Issues Management
- Linear Workspace Integration
- Nicht-technische User Interface
Lernkurve: 1 Tag
Harness Vergleichstabelle
| Kriterium | Everything | Framework | Ruflo | Composio | Your Claude |
|---|---|---|---|---|---|
| Agents (parallel) | 50+ | 3-5 | 100+ | Beliebig | 1-3 |
| Komplexität | Hoch | Niedrig | Sehr Hoch | Mittel | Mittel |
| Setup Time | 3 Tage | 4h | 1-2 Tage | 1 Tag | 1-2 Tage |
| Production Ready | Ja | Mit Tweaks | Ja | Ja | Ja |
| Slack Integration | Nein | Nein | Optional | Nein | Ja |
| GitHub Integration | Stark | Mittel | Mittel | Stark | Stark |
| Monitoring | Minimal | Nein | Umfangreich | Einfach | Minimal |
| RAG Support | Nein | Nein | Ja | Ja | Nein |
| Beste Für | Scale | MVP | Enterprise | Developer | Team Collaboration |
Architektur-Entscheidung: Welcher Harness?
Entscheidungsbaum:
START
↓
Wie viele Agents brauchst du?
├─ 1-2 Agents → Your Claude Engineer oder Framework
├─ 3-10 Agents → Everything Claude Code oder Composio
└─ 50+ Agents → Ruflo
↓
Parallel oder Sequential?
├─ Sequential → Everything Claude Code
└─ Parallel → Composio oder Ruflo
↓
Brauchst du Knowledge Base / RAG?
├─ Ja → Ruflo oder Composio
└─ Nein → Everything Claude Code oder Framework
↓
Wer sind die User?
├─ Entwickler → Composio oder GitHub Integration
├─ Nicht-technisch → Your Claude Engineer + Slack
└─ System-Admin → Ruflo mit Monitoring
Installation Quick-Start
Scenario 1: MVP (schnell starten)
# Claude Agentic Framework
git clone https://github.com/dralgorhythm/claude-agentic-framework.git
cd claude-agentic-framework
npm install
npm run start
# ✓ In 30 min: 3 Agents working
Scenario 2: Production (skalierbar)
# Everything Claude Code
git clone https://github.com/affaan-m/everything-claude-code.git
cd everything-claude-code
npm install
node index.js --config config.prod.json
# ✓ In 3 Tagen: 50+ Agents ready for Enterprise
Scenario 3: Team Collaboration (Slack-first)
# Your Claude Engineer
git clone https://github.com/coleam00/your-claude-engineer.git
cd your-claude-engineer
export SLACK_TOKEN=...
export GITHUB_TOKEN=...
npm run start
# ✓ In 1 Tag: Team kann via Slack arbeiten
Common Patterns
Pattern 1: Analyst → Builder → Reviewer
// Everything Claude Code Config
const agents = [
{
name: "analyst",
role: "Analyse Requirements & Design",
skills: ["code-analysis", "architecture-review"]
},
{
name: "builder",
role: "Implementiert Code",
skills: ["code-generation", "testing"]
},
{
name: "reviewer",
role: "Code Review & QA",
skills: ["code-review", "security-scan"]
}
];
// Workflow: analyst → builder → reviewer
Pattern 2: Parallel Workers + Coordinator
# Composio Framework
agents = [
ParallelAgent("worker-1", tasks["module-a"]),
ParallelAgent("worker-2", tasks["module-b"]),
ParallelAgent("worker-3", tasks["module-c"]),
SequentialAgent("coordinator", ["merge", "test", "deploy"])
]
Pattern 3: Slack-Driven
// Your Claude Engineer
slack.onMessage("@claude build feature X", () => {
agent.run("Build feature X");
// Posts to GitHub issues
// Responds in Slack
});
Häufige Fehler
Fehler 1: Zu viele Agents für falschem Harness
- Problem: 50 Agents mit Framework = Chaos
- Lösung: Everything Claude Code oder Ruflo für Scale
Fehler 2: Kein State Management
- Problem: Agents arbeiten aneinander vorbei
- Lösung: Memory System wie in Everything Claude Code
Fehler 3: Credentials hardcoded
- Problem: Git enthält API Keys
- Lösung: Vault Integration (alle Harnesses unterstützen)
Monitoring & Debugging
Jeder Harness hat unterschiedliche Debugging-Tools:
Everything Claude Code:
node index.js --debug
# → Logs: Agent State, Memory, Tool Calls
Ruflo:
Dashboard: http://localhost:3000
→ Real-Time Agent Status, Resource Usage
Composio:
composio status
# → Parallel Agents, PR Status, Test Results
Quellen und Links
- https://github.com/affaan-m/everything-claude-code
- https://github.com/dralgorhythm/claude-agentic-framework
- https://github.com/ruvnet/ruflo
- https://github.com/ComposioHQ/agent-orchestrator
- https://github.com/coleam00/your-claude-engineer
- https://github.com/wshobson/agents
- https://anthropic.com/docs/claude-code
