{"title":"Coordination multi-agent","slug":"multi-agent-coordination","category":"guides","summary":"Coordonnez plusieurs agents LLM utilisant des minds, tâches et chat Synapse partagés.","audience":["human","llm"],"tags":["guide","multi-agent","coordination","patterns"],"difficulty":"advanced","updated":"2026-06-27","word_count":294,"read_minutes":1,"lang":"fr","translated":true,"requested_lang":"fr","content_markdown":"\n# Coordination multi-agent\n\nQuand vous avez plusieurs agents LLM travaillant sur des tâches liées, Synapse fournit\nla couche de coordination — mémoire partagée, affectation de tâches et chat asynchrone.\n\n## Schémas\n\n### Schéma 1 : Mind partagé (source unique de vérité)\n\nTous les agents partagent une Mind Key. Ils lisent/écrivent le même magasin de mémoire.\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**Cas d'usage :** petite équipe d'agents travaillant sur un projet.\n\n**Configuration :**\n\n```bash\n# Tous les agents utilisent la même Mind Key\nexport SYNAPSE_MIND_KEY=mk_shared_key...\n```\n\n**Coordination via tâches :**\n\n```python\n# L'agent A crée une tâche\ncreate_task(\"Review PR #42\", priority=\"high\")\n\n# L'agent B la récupère\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### Schéma 2 : Minds spécialisés (contextes isolés)\n\nChaque agent a son propre mind. Ils communiquent via un mind de « coordination »\npartagé.\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**Cas d'usage :** agents avec différentes spécialités (codage, revue, déploiement).\n\n**Configuration :**\n\n```bash\n# Agent codeur\nSYNAPSE_MIND_KEY=mk_coder... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Agent relecteur\nSYNAPSE_MIND_KEY=mk_reviewer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n# Agent déploiement\nSYNAPSE_MIND_KEY=mk_deployer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n```\n\n**Coordination via mind partagé :**\n\n```python\n# Le codeur stocke « prêt pour relecture »\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# Le relecteur récupère les demandes de revue\nr = requests.get(f\"{URL}/memory/search?q=ready+for+review\",\n    headers={\"Authorization\": f\"Bearer {COORDINATION_KEY}\"})\n```\n\n### Schéma 3 : hub-and-spoke (orchestrateur)\n\nUn agent orchestrateur central affecte des tâches à des agents travailleurs.\n\n```\n        ┌──────────────┐\n        │ Orchestrator │\n        │    Agent     │\n        └──────┬───────┘\n               │\n    ┌──────────┼──────────┐\n    ▼          ▼          ▼\n┌──────┐  ┌──────┐  ┌──────┐\n│Worker│  │Worker│  │Worker│\n│  A   │  │  B   │  │  C   │\n└──────┘  └──────┘  └──────┘\n```\n\n**Cas d'usage :** workflows complexes avec travail parallèle.\n\n**Implémentation :**\n\n```python\n# Orchestrateur\nclass Orchestrator:\n    def assign_task(self, worker_id, task_description):\n        # Stocker la tâche dans le mind du travailleur (ou mind de coordination partagé)\n        create_task(task_description, priority=\"high\")\n        # Notifier le travailleur 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# Les travailleurs pollinent les tâches affectées\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## Coordination via chat\n\nLes agents peuvent communiquer via le système de chat :\n\n```python\n# L'agent A envoie à l'agent B\nreply(\"@agent-b: Can you review my PR?\")\n\n# L'agent B polline et répond\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> Les messages de chat sont marqués par rôle. Définissez role=agent pour les messages\n> agent-à-agent, role=human pour humain-à-agent.\n\n## Coordination via variables\n\nUtilisez des variables pour une coordination légère (verrous, drapeaux) :\n\n```python\n# Acquérir un verrou\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# Utilisation\nif acquire_lock(\"deploy\"):\n    try:\n        deploy_to_production()\n    finally:\n        release_lock(\"deploy\")\n```\n\n## Bonnes pratiques\n\n> [!TIP]\n> - **Utilisez des minds séparés pour des préoccupations séparées** — ne mélangez pas mémoire de codeur et de relecteur\n> - **Taguez les agents dans le chat** — `@agent-name` pour une adresse claire\n> - **Utilisez des tâches pour l'affectation de travail** — pas le chat (le chat est pour la discussion)\n> - **Implémentez l'idempotence** — les agents peuvent réessayer des opérations échouées\n> - **Journalisez tout** — stockez les décisions en mémoire pour l'auditabilité\n\n## Prochaines étapes\n\n- [Agent LLM persistant](/docs/guides/persistent-llm-agent)\n- [Cookbook LLM](/docs/llm-cookbook/session-start-pattern)\n- [Automatisation par webhooks](/docs/guides/webhook-automation)\n","content_html":"<h1>Coordination multi-agent</h1>\n<p>Quand vous avez plusieurs agents LLM travaillant sur des tâches liées, Synapse fournit\nla couche de coordination — mémoire partagée, affectation de tâches et chat asynchrone.</p>\n<h2>Schémas</h2>\n<h3>Schéma 1 : Mind partagé (source unique de vérité)</h3>\n<p>Tous les agents partagent une Mind Key. Ils lisent/écrivent le même magasin de mémoire.</p>\n<pre><code class=\"hljs language-plaintext\">┌──────────┐  ┌──────────┐  ┌──────────┐\n│ Agent A  │  │ Agent B  │  │ Agent C  │\n└────┬─────┘  └────┬─────┘  └────┬─────┘\n     │             │             │\n     └─────────────┼─────────────┘\n                   ▼\n           ┌──────────────┐\n           │ Shared Mind  │\n           │  (one key)   │\n           └──────────────┘</code></pre><p><strong>Cas d&#39;usage :</strong> petite équipe d&#39;agents travaillant sur un projet.</p>\n<p><strong>Configuration :</strong></p>\n<pre><code class=\"hljs language-bash\"><span class=\"hljs-comment\"># Tous les agents utilisent la même Mind Key</span>\n<span class=\"hljs-built_in\">export</span> SYNAPSE_MIND_KEY=mk_shared_key...</code></pre><p><strong>Coordination via tâches :</strong></p>\n<pre><code class=\"hljs language-python\"><span class=\"hljs-comment\"># L&#x27;agent A crée une tâche</span>\ncreate_task(<span class=\"hljs-string\">&quot;Review PR #42&quot;</span>, priority=<span class=\"hljs-string\">&quot;high&quot;</span>)\n\n<span class=\"hljs-comment\"># L&#x27;agent B la récupère</span>\ntasks = list_tasks(status=<span class=\"hljs-string\">&quot;pending&quot;</span>)\n<span class=\"hljs-keyword\">if</span> tasks:\n    task = tasks[<span class=\"hljs-number\">0</span>]\n    update_task(task[<span class=\"hljs-string\">&quot;id&quot;</span>], status=<span class=\"hljs-string\">&quot;in_progress&quot;</span>)\n    <span class=\"hljs-comment\"># ... do work ...</span>\n    update_task(task[<span class=\"hljs-string\">&quot;id&quot;</span>], status=<span class=\"hljs-string\">&quot;done&quot;</span>)</code></pre><h3>Schéma 2 : Minds spécialisés (contextes isolés)</h3>\n<p>Chaque agent a son propre mind. Ils communiquent via un mind de « coordination »\npartagé.</p>\n<pre><code class=\"hljs language-plaintext\">┌──────────┐  ┌──────────┐  ┌──────────┐\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           └──────────────────┘</code></pre><p><strong>Cas d&#39;usage :</strong> agents avec différentes spécialités (codage, revue, déploiement).</p>\n<p><strong>Configuration :</strong></p>\n<pre><code class=\"hljs language-bash\"><span class=\"hljs-comment\"># Agent codeur</span>\nSYNAPSE_MIND_KEY=mk_coder... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n<span class=\"hljs-comment\"># Agent relecteur</span>\nSYNAPSE_MIND_KEY=mk_reviewer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest\n\n<span class=\"hljs-comment\"># Agent déploiement</span>\nSYNAPSE_MIND_KEY=mk_deployer... MCP_TRANSPORT=stdio npx synapse-mcp-api@latest</code></pre><p><strong>Coordination via mind partagé :</strong></p>\n<pre><code class=\"hljs language-python\"><span class=\"hljs-comment\"># Le codeur stocke « prêt pour relecture »</span>\nCOORDINATION_KEY = <span class=\"hljs-string\">&quot;mk_coordination...&quot;</span>\nrequests.post(<span class=\"hljs-string\">f&quot;<span class=\"hljs-subst\">{URL}</span>/memory&quot;</span>,\n    headers={<span class=\"hljs-string\">&quot;Authorization&quot;</span>: <span class=\"hljs-string\">f&quot;Bearer <span class=\"hljs-subst\">{COORDINATION_KEY}</span>&quot;</span>},\n    json={\n        <span class=\"hljs-string\">&quot;category&quot;</span>: <span class=\"hljs-string\">&quot;project&quot;</span>,\n        <span class=\"hljs-string\">&quot;key&quot;</span>: <span class=\"hljs-string\">&quot;pr_42_ready&quot;</span>,\n        <span class=\"hljs-string\">&quot;content&quot;</span>: <span class=\"hljs-string\">&quot;PR #42 is ready for review. Branch: feature/docs-system&quot;</span>,\n        <span class=\"hljs-string\">&quot;tags&quot;</span>: [<span class=\"hljs-string\">&quot;review&quot;</span>, <span class=\"hljs-string\">&quot;pr-42&quot;</span>],\n        <span class=\"hljs-string\">&quot;priority&quot;</span>: <span class=\"hljs-string\">&quot;high&quot;</span>\n    })\n\n<span class=\"hljs-comment\"># Le relecteur récupère les demandes de revue</span>\nr = requests.get(<span class=\"hljs-string\">f&quot;<span class=\"hljs-subst\">{URL}</span>/memory/search?q=ready+for+review&quot;</span>,\n    headers={<span class=\"hljs-string\">&quot;Authorization&quot;</span>: <span class=\"hljs-string\">f&quot;Bearer <span class=\"hljs-subst\">{COORDINATION_KEY}</span>&quot;</span>})</code></pre><h3>Schéma 3 : hub-and-spoke (orchestrateur)</h3>\n<p>Un agent orchestrateur central affecte des tâches à des agents travailleurs.</p>\n<pre><code class=\"hljs language-plaintext\">        ┌──────────────┐\n        │ Orchestrator │\n        │    Agent     │\n        └──────┬───────┘\n               │\n    ┌──────────┼──────────┐\n    ▼          ▼          ▼\n┌──────┐  ┌──────┐  ┌──────┐\n│Worker│  │Worker│  │Worker│\n│  A   │  │  B   │  │  C   │\n└──────┘  └──────┘  └──────┘</code></pre><p><strong>Cas d&#39;usage :</strong> workflows complexes avec travail parallèle.</p>\n<p><strong>Implémentation :</strong></p>\n<pre><code class=\"hljs language-python\"><span class=\"hljs-comment\"># Orchestrateur</span>\n<span class=\"hljs-keyword\">class</span> <span class=\"hljs-title class_\">Orchestrator</span>:\n    <span class=\"hljs-keyword\">def</span> <span class=\"hljs-title function_\">assign_task</span>(<span class=\"hljs-params\">self, worker_id, task_description</span>):\n        <span class=\"hljs-comment\"># Stocker la tâche dans le mind du travailleur (ou mind de coordination partagé)</span>\n        create_task(task_description, priority=<span class=\"hljs-string\">&quot;high&quot;</span>)\n        <span class=\"hljs-comment\"># Notifier le travailleur via chat</span>\n        reply(<span class=\"hljs-string\">f&quot;@<span class=\"hljs-subst\">{worker_id}</span>: New task — <span class=\"hljs-subst\">{task_description}</span>&quot;</span>)\n    \n    <span class=\"hljs-keyword\">def</span> <span class=\"hljs-title function_\">check_progress</span>(<span class=\"hljs-params\">self</span>):\n        tasks = list_tasks(status=<span class=\"hljs-string\">&quot;in_progress&quot;</span>)\n        <span class=\"hljs-keyword\">for</span> t <span class=\"hljs-keyword\">in</span> tasks:\n            <span class=\"hljs-built_in\">print</span>(<span class=\"hljs-string\">f&quot;<span class=\"hljs-subst\">{t[<span class=\"hljs-string\">&#x27;title&#x27;</span>]}</span>: <span class=\"hljs-subst\">{t[<span class=\"hljs-string\">&#x27;status&#x27;</span>]}</span>&quot;</span>)\n\n<span class=\"hljs-comment\"># Les travailleurs pollinent les tâches affectées</span>\n<span class=\"hljs-keyword\">class</span> <span class=\"hljs-title class_\">Worker</span>:\n    <span class=\"hljs-keyword\">def</span> <span class=\"hljs-title function_\">run</span>(<span class=\"hljs-params\">self</span>):\n        <span class=\"hljs-keyword\">while</span> <span class=\"hljs-literal\">True</span>:\n            tasks = list_tasks(status=<span class=\"hljs-string\">&quot;pending&quot;</span>)\n            <span class=\"hljs-keyword\">for</span> t <span class=\"hljs-keyword\">in</span> tasks:\n                <span class=\"hljs-keyword\">if</span> assigned_to_me(t):\n                    update_task(t[<span class=\"hljs-string\">&quot;id&quot;</span>], status=<span class=\"hljs-string\">&quot;in_progress&quot;</span>)\n                    result = do_work(t)\n                    update_task(t[<span class=\"hljs-string\">&quot;id&quot;</span>], status=<span class=\"hljs-string\">&quot;done&quot;</span>)\n                    reply(<span class=\"hljs-string\">f&quot;Completed: <span class=\"hljs-subst\">{t[<span class=\"hljs-string\">&#x27;title&#x27;</span>]}</span>&quot;</span>)\n            time.sleep(<span class=\"hljs-number\">60</span>)</code></pre><h2>Coordination via chat</h2>\n<p>Les agents peuvent communiquer via le système de chat :</p>\n<pre><code class=\"hljs language-python\"><span class=\"hljs-comment\"># L&#x27;agent A envoie à l&#x27;agent B</span>\nreply(<span class=\"hljs-string\">&quot;@agent-b: Can you review my PR?&quot;</span>)\n\n<span class=\"hljs-comment\"># L&#x27;agent B polline et répond</span>\n<span class=\"hljs-keyword\">for</span> msg <span class=\"hljs-keyword\">in</span> poll_messages():\n    <span class=\"hljs-keyword\">if</span> <span class=\"hljs-string\">&quot;@agent-b&quot;</span> <span class=\"hljs-keyword\">in</span> msg[<span class=\"hljs-string\">&quot;content&quot;</span>]:\n        reply(<span class=\"hljs-string\">f&quot;@agent-a: Sure, looking at it now.&quot;</span>)</code></pre><div class=\"callout callout-note\">Les messages de chat sont marqués par rôle. Définissez role=agent pour les messages\nagent-à-agent, role=human pour humain-à-agent.</div><h2>Coordination via variables</h2>\n<p>Utilisez des variables pour une coordination légère (verrous, drapeaux) :</p>\n<pre><code class=\"hljs language-python\"><span class=\"hljs-comment\"># Acquérir un verrou</span>\n<span class=\"hljs-keyword\">def</span> <span class=\"hljs-title function_\">acquire_lock</span>(<span class=\"hljs-params\">name</span>):\n    r = requests.post(<span class=\"hljs-string\">f&quot;<span class=\"hljs-subst\">{URL}</span>/var&quot;</span>,\n        headers={<span class=\"hljs-string\">&quot;Authorization&quot;</span>: <span class=\"hljs-string\">f&quot;Bearer <span class=\"hljs-subst\">{KEY}</span>&quot;</span>},\n        json={<span class=\"hljs-string\">&quot;key&quot;</span>: <span class=\"hljs-string\">f&quot;lock_<span class=\"hljs-subst\">{name}</span>&quot;</span>, <span class=\"hljs-string\">&quot;value&quot;</span>: <span class=\"hljs-string\">&quot;acquired&quot;</span>})\n    <span class=\"hljs-keyword\">return</span> <span class=\"hljs-literal\">True</span>\n\n<span class=\"hljs-keyword\">def</span> <span class=\"hljs-title function_\">release_lock</span>(<span class=\"hljs-params\">name</span>):\n    requests.delete(<span class=\"hljs-string\">f&quot;<span class=\"hljs-subst\">{URL}</span>/var/lock_<span class=\"hljs-subst\">{name}</span>&quot;</span>,\n        headers={<span class=\"hljs-string\">&quot;Authorization&quot;</span>: <span class=\"hljs-string\">f&quot;Bearer <span class=\"hljs-subst\">{KEY}</span>&quot;</span>})\n\n<span class=\"hljs-comment\"># Utilisation</span>\n<span class=\"hljs-keyword\">if</span> acquire_lock(<span class=\"hljs-string\">&quot;deploy&quot;</span>):\n    <span class=\"hljs-keyword\">try</span>:\n        deploy_to_production()\n    <span class=\"hljs-keyword\">finally</span>:\n        release_lock(<span class=\"hljs-string\">&quot;deploy&quot;</span>)</code></pre><h2>Bonnes pratiques</h2>\n<div class=\"callout callout-ok\"></div><h2>Prochaines étapes</h2>\n<ul>\n<li><a href=\"/docs/guides/persistent-llm-agent\">Agent LLM persistant</a></li>\n<li><a href=\"/docs/llm-cookbook/session-start-pattern\">Cookbook LLM</a></li>\n<li><a href=\"/docs/guides/webhook-automation\">Automatisation par webhooks</a></li>\n</ul>\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"]}