agentleFS
khalilbenazkhalilbenaz/claude-skills-collection/skills/agent-a2a-protocol-guide/SKILL.md

agent-a2a-protocol-guide

Skill28 starsChanged 14 days agoMITView on GitHub

Guide du protocole Agent-to-Agent (A2A) de Google pour l'interopérabilité entre agents IA — découverte, communication et collaboration inter-agents, avec exemples de code, critères de décision et pièges à éviter. Se déclenche avec "A2A", "agent-to-agent", "protocole A2A", "Google A2A", "interopérabilité agents". Also triggers on "agent-to-agent protocol", "agent interoperability", "connect two agents".

  • Installs packages

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never say "happy to help", "sure!", "great question", "let me", or similar.

    line 248
  2. Distinctive:

    Stop when done. No summary, no recap, no trailing commentary.

    line 251
  3. Distinctive:

    No politeness wrappers. Direct and blunt.

    line 252
  4. Distinctive:

    No emoji unless asked.

    line 255

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Guide du Protocole A2A

Concepts fondamentaux

ComposantRôle
Agent CardFichier JSON décrivant l'agent (skills, endpoint, auth) — publié sur /.well-known/agent.json
TaskUnité de travail délégué ; cycle : submitted → working → completed / failed / canceled
ArtifactRésultat produit par une tâche (fichier, JSON, message)
PartFragment d'un message ou artifact : TextPart, FilePart, DataPart
Client / ServerUn agent peut être les deux selon la direction de l'appel

Workflow en étapes

1. Définir l'Agent Card

Fichier minimal exposé sur GET /.well-known/agent.json :

{
  "name": "DataExtractorAgent",
  "description": "Extrait des données structurées depuis des documents PDF",
  "url": "https://agent.example.com/a2a",
  "version": "1.0.0",
  "capabilities": {
    "streaming": true,
    "pushNotifications": false
  },
  "authentication": {
    "schemes": ["Bearer"]
  },
  "skills": [
    {
      "id": "extract-invoice",
      "name": "Invoice Extraction",
      "description": "Retourne les champs clés d'une facture PDF",
      "examples": ["Extrais les montants de cette facture"],
      "inputModes": ["application/pdf"],
      "outputModes": ["application/json"]
    }
  ]
}

Critère de décision — streaming : active streaming: true si les tâches durent > 5 s ou produisent des résultats progressifs. Sinon, mode synchrone suffit.


2. Implémenter le serveur A2A (JSON-RPC 2.0)

Endpoint unique POST /a2a — méthodes requises :

MéthodeUsage
tasks/sendTâche synchrone (réponse directe)
tasks/sendSubscribeTâche longue — SSE stream
tasks/getConsulter statut + artifacts
tasks/cancelAnnuler une tâche en cours
# FastAPI — squelette minimal
from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse, StreamingResponse
import uuid, asyncio

app = FastAPI()

tasks_store = {}  # En prod : Redis ou DB

@app.post("/a2a")
async def a2a_handler(request: Request):
    body = await request.json()
    method = body.get("method")
    params = body.get("params", {})
    rpc_id = body.get("id")

    if method == "tasks/send":
        task_id = str(uuid.uuid4())
        result = await run_task(params)
        tasks_store[task_id] = {"status": "completed", "result": result}
        return JSONResponse({
            "jsonrpc": "2.0", "id": rpc_id,
            "result": {"id": task_id, "status": {"state": "completed"}, "artifacts": [result]}
        })

    if method == "tasks/get":
        task_id = params["id"]
        task = tasks_store.get(task_id)
        if not task:
            return JSONResponse({"jsonrpc": "2.0", "id": rpc_id,
                                 "error": {"code": -32001, "message": "Task not found"}})
        return JSONResponse({"jsonrpc": "2.0", "id": rpc_id, "result": task})

    return JSONResponse({"jsonrpc": "2.0", "id": rpc_id,
                         "error": {"code": -32601, "message": "Method not found"}})

3. Appeler un agent distant (côté client)

import httpx, json

async def send_task(agent_url: str, message: str, token: str) -> dict:
    payload = {
        "jsonrpc": "2.0", "id": "1", "method": "tasks/send",
        "params": {
            "message": {
                "role": "user",
                "parts": [{"type": "text", "text": message}]
            }
        }
    }
    async with httpx.AsyncClient() as client:
        resp = await client.post(
            agent_url,
            json=payload,
            headers={"Authorization": f"Bearer {token}"},
            timeout=30
        )
        resp.raise_for_status()
        return resp.json()["result"]

4. Streaming SSE (tasks/sendSubscribe)

from fastapi.responses import StreamingResponse

@app.post("/a2a/stream")
async def stream_task(request: Request):
    body = await request.json()

    async def event_generator():
        # Mise à jour intermédiaire
        yield f"data: {json.dumps({'type': 'TaskStatusUpdateEvent', 'status': {'state': 'working'}})}\n\n"
        await asyncio.sleep(1)
        # Résultat final
        artifact = {"type": "DataPart", "data": {"key": "value"}}
        yield f"data: {json.dumps({'type': 'TaskArtifactUpdateEvent', 'artifact': artifact})}\n\n"
        yield f"data: {json.dumps({'type': 'TaskStatusUpdateEvent', 'status': {'state': 'completed'}})}\n\n"

    return StreamingResponse(event_generator(), media_type="text/event-stream")

5. Découverte et sélection d'agent

async def discover_agent(base_url: str) -> dict:
    async with httpx.AsyncClient() as client:
        r = await client.get(f"{base_url}/.well-known/agent.json", timeout=5)
        r.raise_for_status()
        return r.json()

# Critère de sélection
def select_agent(agents: list[dict], required_skill: str) -> dict | None:
    for a in agents:
        skills = [s["id"] for s in a.get("skills", [])]
        if required_skill in skills:
            return a
    return None

Registre simple : maintenir une liste d'URLs connues + cache de leurs Agent Cards (TTL 5 min). Pour du dynamique, utiliser un service de registre central (ex. : endpoint GET /agents dans l'orchestrateur).


6. Sécurité

  • Transport : HTTPS obligatoire ; rejeter toute requête HTTP plain.
  • Auth : OAuth 2.0 Bearer token ou API key dans Authorization header — déclarer le scheme dans l'Agent Card.
  • Validation : vérifier la signature de l'Agent Card si elle provient d'un registre tiers.
  • Autorisation : liste blanche d'agents autorisés par skill — refuser les appels d'agents inconnus.
# Middleware auth minimal
from fastapi import Depends, HTTPException, Header

async def verify_token(authorization: str = Header(...)):
    if not authorization.startswith("Bearer "):
        raise HTTPException(status_code=401)
    token = authorization[7:]
    if token not in ALLOWED_TOKENS:
        raise HTTPException(status_code=403)

7. Orchestration multi-agents

Pattern fan-out / fan-in :

async def orchestrate(agents: list[str], prompt: str) -> list:
    tasks = [send_task(url, prompt, TOKEN) for url in agents]
    results = await asyncio.gather(*tasks, return_exceptions=True)
    return [r for r in results if not isinstance(r, Exception)]

Pattern chaîne séquentielle : sortie artifact de l'agent N devient le message d'entrée de l'agent N+1. Extraire le bon Part depuis artifacts[0].parts.


8. Monitoring et debug

  • Propager un X-Correlation-ID dans tous les appels inter-agents.
  • Logger : task_id, agent_url, method, status_transition, duration_ms.
  • Alertes : tâches en état working depuis > timeout configuré → déclencher tasks/cancel.
  • Sanity check périodique : GET /.well-known/agent.json sur chaque agent du registre.

Garde-fous et anti-patterns

Anti-patternProblèmeCorrect
Stocker l'état des tâches en mémoire processPerdu au redémarrageRedis / DB persistante
Pas de timeout sur tasks/sendBlocage indéfiniTimeout HTTP + tasks/cancel si dépassé
Agent Card statique en cache illimitéCapabilities obsolètesTTL max 5 min, invalider sur erreur 404/401
Exposer /.well-known/agent.json sans authEnumération par tiersAuth optionnelle mais recommandée en prod interne
Format propriétaire dans les partsInterop casséeUtiliser TextPart, FilePart, DataPart standard
Ignorer les erreurs return_exceptions en gatherRésultats silencieusement manquantsLogger chaque exception, circuit breaker si taux > seuil

Bonnes pratiques 2026

  • Versionner l'Agent Card ("version") et gérer la compatibilité ascendante des skills.
  • Implémenter idempotency via un taskId côté client — le serveur doit dédupliquer les renvois.
  • Limiter la taille des artifacts inline : au-delà de 1 Mo, retourner une URL signée (FilePart avec uri).
  • Déclarer les inputModes / outputModes MIME précis : facilite la sélection automatique d'agent.
  • Tester la conformité avec le SDK officiel Google A2A (pip install a2a-sdk ou npm install @google/a2a).

Communication Rules — MANDATORY

  • Ultra-concise. No filler, no preamble, no pleasantries.
  • Never say "happy to help", "sure!", "great question", "let me", or similar.
  • Tool first, talk second. Act before explaining.
  • Result first. Lead with outcome, not process.
  • Stop when done. No summary, no recap, no trailing commentary.
  • No politeness wrappers. Direct and blunt.
  • Minimum words. If one word works, do not use ten.
  • No unsolicited explanations.
  • No emoji unless asked.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.