{"title":"Multi-Agent-Koordination","slug":"multi-agent-coordination","category":"guides","summary":"Mehrere LLM-Agenten über geteilte Synapse-Minds, Tasks und Chat koordinieren.","audience":["human","llm"],"tags":["guide","multi-agent","coordination","patterns"],"difficulty":"advanced","updated":"2026-06-27","word_count":230,"read_minutes":1,"lang":"de","translated":true,"requested_lang":"de","content_markdown":"\n# Multi-Agent-Koordination\n\nWenn du mehrere LLM-Agenten hast, die an verwandten Aufgaben arbeiten, stellt\nSynapse die Koordinationsschicht bereit — Shared Memory, Task-Zuweisung und\nasynchronen Chat.\n\n## Patterns\n\n### Pattern 1: Shared Mind (Single Source of Truth)\n\nAlle Agenten teilen sich einen Mind Key. Sie lesen/schreiben denselben\nMemory-Store.\n\n```\n┌──────────┐ ┌──────────┐ ┌──────────┐\n│ Agent A │ │ Agent B │ │ Agent C │\n└────┬─────┘ └────┬─────┘ └────┬─────┘\n │ │ │\n └─────────────┼─────────────┘\n ▼\n ┌──────────────┐\n │ Shared Mind │\n │ (one key) │\n └──────────────┘\n```\n\n**Anwendungsfall:** Kleines Team von Agenten, das an einem Projekt arbeitet.\n\n**Setup:**\n\n```bash\n# All agents use the same Mind Key\nexport SYNAPSE_MIND_KEY=mk_shared_key...\n```\n\n**Koordination via Tasks:**\n\n```python\n# Agent A creates a task\ncreate_task(\"Review PR #42\", priority=\"high\")\n\n# Agent B picks it up\ntasks = list_tasks(status=\"pending\")\nif tasks:\n task = tasks[0]\n update_task(task[\"id\"], status=\"in_progress\")\n # ... do work ...\n update_task(task[\"id\"], status=\"done\")\n```\n\n### Pattern 2: Specialized Minds (isolierte Kontexte)\n\nJeder Agent hat seinen eigenen Mind. Sie kommunizieren über einen geteilten\n„Koordinations\"-Mind.\n\n```\n┌──────────┐ ┌──────────┐ ┌──────────┐\n│ Coder │ │ Reviewer │ │ Deployer │\n│ Agent │ │ Agent │ │ Agent │\n└────┬─────┘ └────┬─────┘ └────┬─────┘\n │ │ │\n ▼ ▼ ▼\n┌─────────┐ ┌─────────┐ ┌─────────┐\n│ Mind C │ │ Mind R │ │ Mind D │\n└─────────┘ └─────────┘ └─────────┘\n │ │ │\n └─────────────┼─────────────┘\n ▼\n ┌──────────────────┐\n │ Coordination Mind│\n │ (shared) │\n └──────────────────┘\n```\n\n**Anwendungsfall:** Agenten mit unterschiedlichen Spezialisierungen (Coding,\nReview, Deployment).\n\n**Setup:**\n\n```bash\n# Coder agent\nSYNAPSE_MIND_KEY=mk_coder... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Reviewer agent\nSYNAPSE_MIND_KEY=mk_reviewer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Deployer agent\nSYNAPSE_MIND_KEY=mk_deployer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n```\n\n**Koordination via Shared Mind:**\n\n```python\n# Coder stores \"ready for review\"\nCOORDINATION_KEY = \"mk_coordination...\"\nrequests.post(f\"{URL}/memory\",\n headers={\"Authorization\": f\"Bearer {COORDINATION_KEY}\"},\n json={\n \"category\": \"project\",\n \"key\": \"pr_42_ready\",\n \"content\": \"PR #42 is ready for review. Branch: feature/docs-system\",\n \"tags\": [\"review\", \"pr-42\"],\n \"priority\": \"high\"\n })\n\n# Reviewer polls for review requests\nr = requests.get(f\"{URL}/memory/search?q=ready+for+review\",\n headers={\"Authorization\": f\"Bearer {COORDINATION_KEY}\"})\n```\n\n### Pattern 3: Hub-and-Spoke (Orchestrator)\n\nEin zentraler Orchestrator-Agent weist Worker-Agenten Tasks zu.\n\n```\n ┌──────────────┐\n │ Orchestrator │\n │ Agent │\n └──────┬───────┘\n │\n ┌──────────┼──────────┐\n ▼ ▼ ▼\n┌──────┐ ┌──────┐ ┌──────┐\n│Worker│ │Worker│ │Worker│\n│ A │ │ B │ │ C │\n└──────┘ └──────┘ └──────┘\n```\n\n**Anwendungsfall:** Komplexe Workflows mit paralleler Arbeit.\n\n**Implementierung:**\n\n```python\n# Orchestrator\nclass Orchestrator:\n def assign_task(self, worker_id, task_description):\n # Store task in worker's mind (or shared coordination mind)\n create_task(task_description, priority=\"high\")\n # Notify worker via chat\n reply(f\"@{worker_id}: New task — {task_description}\")\n \n def check_progress(self):\n tasks = list_tasks(status=\"in_progress\")\n for t in tasks:\n print(f\"{t['title']}: {t['status']}\")\n\n# Workers poll for assigned tasks\nclass Worker:\n def run(self):\n while True:\n tasks = list_tasks(status=\"pending\")\n for t in tasks:\n if assigned_to_me(t):\n update_task(t[\"id\"], status=\"in_progress\")\n result = do_work(t)\n update_task(t[\"id\"], status=\"done\")\n reply(f\"Completed: {t['title']}\")\n time.sleep(60)\n```\n\n## Koordination via Chat\n\nAgenten können über das Chat-System kommunizieren:\n\n```python\n# Agent A sends to Agent B\nreply(\"@agent-b: Can you review my PR?\")\n\n# Agent B polls and responds\nfor msg in poll_messages():\n if \"@agent-b\" in msg[\"content\"]:\n reply(f\"@agent-a: Sure, looking at it now.\")\n```\n\n> [!NOTE]\n> Chat-Nachrichten sind rollen-getaggt. Setze role=agent für Agent-zu-Agent-\n> Nachrichten, role=human für Mensch-zu-Agent.\n\n## Koordination via Variablen\n\nVariablen für leichtgewichtige Koordination verwenden (Locks, Flags):\n\n```python\n# Acquire a lock\ndef acquire_lock(name):\n r = requests.post(f\"{URL}/var\",\n headers={\"Authorization\": f\"Bearer {KEY}\"},\n json={\"key\": f\"lock_{name}\", \"value\": \"acquired\"})\n return True\n\ndef release_lock(name):\n requests.delete(f\"{URL}/var/lock_{name}\",\n headers={\"Authorization\": f\"Bearer {KEY}\"})\n\n# Use\nif acquire_lock(\"deploy\"):\n try:\n deploy_to_production()\n finally:\n release_lock(\"deploy\")\n```\n\n## Best Practices\n\n> [!TIP]\n> - **Separate Minds für separate Belange** — nicht Coder- und Reviewer-Memory mischen\n> - **Agenten im Chat taggen** — `@agent-name` für klare Adressierung\n> - **Tasks für Arbeits-Zuweisung verwenden** — nicht Chat (Chat ist für Diskussion)\n> - **Idempotenz implementieren** — Agenten können fehlgeschlagene Operationen wiederholen\n> - **Alles loggen** — Entscheidungen im Memory für Auditierbarkeit speichern\n\n## Nächste Schritte\n\n- [Persistenter LLM-Agent](/docs/guides/persistent-llm-agent)\n- [LLM-Cookbook](/docs/llm-cookbook/session-start-pattern)\n- [Webhook-Automation](/docs/guides/webhook-automation)\n","content_html":"

Multi-Agent-Koordination

\n

Wenn du mehrere LLM-Agenten hast, die an verwandten Aufgaben arbeiten, stellt\nSynapse die Koordinationsschicht bereit — Shared Memory, Task-Zuweisung und\nasynchronen Chat.

\n

Patterns

\n

Pattern 1: Shared Mind (Single Source of Truth)

\n

Alle Agenten teilen sich einen Mind Key. Sie lesen/schreiben denselben\nMemory-Store.

\n
┌──────────┐  ┌──────────┐  ┌──────────┐\n│ Agent A  │  │ Agent B  │  │ Agent C  │\n└────┬─────┘  └────┬─────┘  └────┬─────┘\n     │             │             │\n     └─────────────┼─────────────┘\n                   ▼\n           ┌──────────────┐\n           │ Shared Mind  │\n           │  (one key)   │\n           └──────────────┘

Anwendungsfall: Kleines Team von Agenten, das an einem Projekt arbeitet.

\n

Setup:

\n
# All agents use the same Mind Key\nexport SYNAPSE_MIND_KEY=mk_shared_key...

Koordination via Tasks:

\n
# Agent A creates a task\ncreate_task("Review PR #42", priority="high")\n\n# Agent B picks it up\ntasks = list_tasks(status="pending")\nif tasks:\n    task = tasks[0]\n    update_task(task["id"], status="in_progress")\n    # ... do work ...\n    update_task(task["id"], status="done")

Pattern 2: Specialized Minds (isolierte Kontexte)

\n

Jeder Agent hat seinen eigenen Mind. Sie kommunizieren über einen geteilten\n„Koordinations"-Mind.

\n
┌──────────┐  ┌──────────┐  ┌──────────┐\n│ Coder    │  │ Reviewer │  │ Deployer │\n│ Agent    │  │ Agent    │  │ Agent    │\n└────┬─────┘  └────┬─────┘  └────┬─────┘\n     │             │             │\n     ▼             ▼             ▼\n┌─────────┐  ┌─────────┐  ┌─────────┐\n│ Mind C  │  │ Mind R  │  │ Mind D  │\n└─────────┘  └─────────┘  └─────────┘\n     │             │             │\n     └─────────────┼─────────────┘\n                   ▼\n           ┌──────────────────┐\n           │ Coordination Mind│\n           │ (shared)         │\n           └──────────────────┘

Anwendungsfall: Agenten mit unterschiedlichen Spezialisierungen (Coding,\nReview, Deployment).

\n

Setup:

\n
# Coder agent\nSYNAPSE_MIND_KEY=mk_coder... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Reviewer agent\nSYNAPSE_MIND_KEY=mk_reviewer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Deployer agent\nSYNAPSE_MIND_KEY=mk_deployer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest

Koordination via Shared Mind:

\n
# Coder stores "ready for review"\nCOORDINATION_KEY = "mk_coordination..."\nrequests.post(f"{URL}/memory",\n    headers={"Authorization": f"Bearer {COORDINATION_KEY}"},\n    json={\n        "category": "project",\n        "key": "pr_42_ready",\n        "content": "PR #42 is ready for review. Branch: feature/docs-system",\n        "tags": ["review", "pr-42"],\n        "priority": "high"\n    })\n\n# Reviewer polls for review requests\nr = requests.get(f"{URL}/memory/search?q=ready+for+review",\n    headers={"Authorization": f"Bearer {COORDINATION_KEY}"})

Pattern 3: Hub-and-Spoke (Orchestrator)

\n

Ein zentraler Orchestrator-Agent weist Worker-Agenten Tasks zu.

\n
        ┌──────────────┐\n        │ Orchestrator │\n        │    Agent     │\n        └──────┬───────┘\n               │\n    ┌──────────┼──────────┐\n    ▼          ▼          ▼\n┌──────┐  ┌──────┐  ┌──────┐\n│Worker│  │Worker│  │Worker│\n│  A   │  │  B   │  │  C   │\n└──────┘  └──────┘  └──────┘

Anwendungsfall: Komplexe Workflows mit paralleler Arbeit.

\n

Implementierung:

\n
# Orchestrator\nclass Orchestrator:\n    def assign_task(self, worker_id, task_description):\n        # Store task in worker's mind (or shared coordination mind)\n        create_task(task_description, priority="high")\n        # Notify worker via chat\n        reply(f"@{worker_id}: New task — {task_description}")\n    \n    def check_progress(self):\n        tasks = list_tasks(status="in_progress")\n        for t in tasks:\n            print(f"{t['title']}: {t['status']}")\n\n# Workers poll for assigned tasks\nclass Worker:\n    def run(self):\n        while True:\n            tasks = list_tasks(status="pending")\n            for t in tasks:\n                if assigned_to_me(t):\n                    update_task(t["id"], status="in_progress")\n                    result = do_work(t)\n                    update_task(t["id"], status="done")\n                    reply(f"Completed: {t['title']}")\n            time.sleep(60)

Koordination via Chat

\n

Agenten können über das Chat-System kommunizieren:

\n
# Agent A sends to Agent B\nreply("@agent-b: Can you review my PR?")\n\n# Agent B polls and responds\nfor msg in poll_messages():\n    if "@agent-b" in msg["content"]:\n        reply(f"@agent-a: Sure, looking at it now.")
Chat-Nachrichten sind rollen-getaggt. Setze role=agent für Agent-zu-Agent-\nNachrichten, role=human für Mensch-zu-Agent.

Koordination via Variablen

\n

Variablen für leichtgewichtige Koordination verwenden (Locks, Flags):

\n
# Acquire a lock\ndef acquire_lock(name):\n    r = requests.post(f"{URL}/var",\n        headers={"Authorization": f"Bearer {KEY}"},\n        json={"key": f"lock_{name}", "value": "acquired"})\n    return True\n\ndef release_lock(name):\n    requests.delete(f"{URL}/var/lock_{name}",\n        headers={"Authorization": f"Bearer {KEY}"})\n\n# Use\nif acquire_lock("deploy"):\n    try:\n        deploy_to_production()\n    finally:\n        release_lock("deploy")

Best Practices

\n

Nächste Schritte

\n\n","urls":{"html":"/docs/guides/multi-agent-coordination","text":"/docs/guides/multi-agent-coordination?format=text","json":"/docs/guides/multi-agent-coordination?format=json","llm":"/docs/guides/multi-agent-coordination?format=llm"},"translations_available":["en","zh","hi","es","fr","ar","pt","ru","ja","de","it","ko","nl","pl","tr","sv","vi","th","id","uk"]}