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

agent-deployment-guide

Skill28 starsChanged 14 days agoMITView on GitHub

Déploiement d'agents IA en production avec scalabilité et fiabilité. Se déclenche avec "déployer agent", "agent en production", "agent API", "hosting agent", "agent scaling", "agent infrastructure", "servir un agent", "agent cloud". Also triggers on "deploy an agent", "agent in production", "agent hosting".

  • Reads credentials
  • 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 292
  2. Distinctive:

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

    line 295
  3. Distinctive:

    No politeness wrappers. Direct and blunt.

    line 296
  4. Distinctive:

    No emoji unless asked.

    line 299

≈ “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

Agent Deployment Guide

Quand utiliser ce skill

Passage d'un agent fonctionnel en local vers un déploiement production fiable et scalable. Couvre API synchrone, worker asynchrone, webhook, tâche planifiée, streaming temps réel. Applicable sur AWS, Azure, GCP, ou infrastructure on-premise.


Étape 1 — Choisir le pattern de déploiement

PatternQuand l'utiliserLatence maxExemple
API synchrone (FastAPI)Usage interactif, réponse < 30s29sChatbot, Q&A
Worker async (Celery/Bull)Tâches longues, batchIllimitéAnalyse de docs, rapport
Webhook handlerÉvénements externes (GitHub, Slack)3s (ACK)Bot Slack, CI/CD agent
Scheduled agent (cron)Récurrent, pas de déclencheur externeN/ARapport hebdo, cleanup
Streaming SSE/WebSocketUX conversationnelle temps réel< 1s TTFBAssistant interactif

Critère décisif : si la réponse prend > 30s → worker async + polling/webhook. Sinon → API synchrone.


Étape 2 — Containeriser l'agent

# Dockerfile multi-stage (léger et sécurisé)
FROM python:3.12-slim AS builder
WORKDIR /build
COPY requirements.txt .
RUN pip install --no-cache-dir --prefix=/install -r requirements.txt

FROM python:3.12-slim
WORKDIR /app
COPY --from=builder /install /usr/local
COPY . .
# Pas de root en prod
RUN useradd -m appuser && chown -R appuser /app
USER appuser
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8080", "--workers", "4"]

Points critiques :

  • Tag Docker immutable (image:v1.2.3), jamais latest en prod
  • Secrets → volume/secret manager, jamais COPY .env dans l'image
  • Model weights → volume monté ou téléchargement S3 au démarrage, pas dans l'image

Étape 3 — Wrapper API robuste

# main.py — FastAPI avec health checks et streaming
from fastapi import FastAPI, HTTPException
from fastapi.responses import StreamingResponse
from pydantic import BaseModel, Field
import asyncio, uuid, os

app = FastAPI()

class AgentRequest(BaseModel):
    message: str = Field(..., max_length=4000)
    conversation_id: str = Field(default_factory=lambda: str(uuid.uuid4()))

@app.get("/health")   # liveness probe
async def health(): return {"status": "ok"}

@app.get("/ready")    # readiness probe
async def ready():
    # Vérifier dépendances critiques
    try:
        await check_llm_api()  # ping minimal
        await check_redis()
    except Exception as e:
        raise HTTPException(503, detail=str(e))
    return {"status": "ready"}

@app.post("/run")
async def run_agent(req: AgentRequest):
    async def stream():
        async for chunk in agent.run_stream(req.message, req.conversation_id):
            yield f"data: {chunk}\n\n"
    return StreamingResponse(stream(), media_type="text/event-stream")

Étape 4 — State management externalisé (obligatoire)

Toute l'état doit vivre hors du processus pour permettre le scaling horizontal.

import redis.asyncio as redis
import json

r = redis.from_url(os.environ["REDIS_URL"])

async def save_thread(conversation_id: str, messages: list):
    await r.setex(
        f"thread:{conversation_id}",
        86400,  # TTL 24h
        json.dumps(messages)
    )

async def load_thread(conversation_id: str) -> list:
    data = await r.get(f"thread:{conversation_id}")
    return json.loads(data) if data else []
  • Redis : sessions courtes, cache tool results (TTL court)
  • PostgreSQL : historique long terme, audit trail, facturation
  • Ne jamais stocker l'état dans une variable globale ou en mémoire d'instance

Étape 5 — Worker async pour tâches longues

# tasks.py — Celery worker
from celery import Celery
app_celery = Celery("agent", broker=os.environ["REDIS_URL"])

@app_celery.task(bind=True, max_retries=3, default_retry_delay=60)
def run_long_agent(self, task_id: str, payload: dict):
    try:
        result = agent.run(payload)
        save_result(task_id, result)
    except Exception as exc:
        self.retry(exc=exc)

# Endpoint de soumission
@app.post("/submit")
async def submit(req: AgentRequest):
    task_id = str(uuid.uuid4())
    run_long_agent.delay(task_id, req.dict())
    return {"task_id": task_id, "status_url": f"/status/{task_id}"}

@app.get("/status/{task_id}")
async def status(task_id: str):
    result = get_result(task_id)  # Redis ou DB
    return result or {"status": "pending"}

Étape 6 — Scaling et résilience

Auto-scaling Kubernetes (HPA) :

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: agent-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: agent
  minReplicas: 2
  maxReplicas: 20
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 60

Circuit breaker avec tenacity :

from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type
import httpx

@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=2, max=10),
    retry=retry_if_exception_type(httpx.HTTPError)
)
async def call_llm(prompt: str) -> str:
    async with httpx.AsyncClient(timeout=25.0) as client:
        response = await client.post(LLM_ENDPOINT, json={"prompt": prompt})
        response.raise_for_status()
        return response.json()["text"]

Étape 7 — Secrets et configuration

# AWS Secrets Manager — récupération au démarrage
aws secretsmanager get-secret-value \
  --secret-id prod/agent/anthropic-key \
  --query SecretString --output text

# Kubernetes Secret (injecter en env, pas en fichier)
kubectl create secret generic agent-secrets \
  --from-literal=ANTHROPIC_API_KEY=sk-...

# Dans le pod spec
envFrom:
  - secretRef:
      name: agent-secrets

Règle : .env uniquement en dev local. En prod → secret manager ou Kubernetes Secrets.


Étape 8 — Rollback et canary release

# Rollback immédiat Kubernetes
kubectl rollout undo deployment/agent
kubectl rollout status deployment/agent

# Canary : 10% vers v2, 90% vers v1 (nginx ingress)
# Annoter l'ingress canary :
kubectl annotate ingress agent-canary \
  nginx.ingress.kubernetes.io/canary="true" \
  nginx.ingress.kubernetes.io/canary-weight="10"

Étape 9 — Observabilité minimale obligatoire (dès J1)

# Prometheus metrics — instrumenter dès le début
from prometheus_client import Counter, Histogram, generate_latest
import time

REQUEST_COUNT = Counter("agent_requests_total", "Total requests", ["status"])
REQUEST_LATENCY = Histogram("agent_request_duration_seconds", "Latency")
LLM_COST = Counter("agent_llm_tokens_total", "Tokens used", ["model"])

@app.middleware("http")
async def metrics_middleware(request, call_next):
    start = time.time()
    response = await call_next(request)
    REQUEST_LATENCY.observe(time.time() - start)
    REQUEST_COUNT.labels(status=response.status_code).inc()
    return response

@app.get("/metrics")
async def metrics():
    return Response(generate_latest(), media_type="text/plain")

Métriques indispensables : latence p50/p95/p99, taux d'erreur, tokens LLM consommés, coût/requête, queue depth.


Garde-fous — Anti-patterns et pièges courants

PiègeSymptômeRemède
État en mémoire d'instanceErreurs aléatoires après scalingExternaliser tout dans Redis/DB
Tag latest en prodRollback impossibleTags immutables v1.2.3
Timeout LLM trop longWorkers bloqués, queue saturéeTimeout 25-28s max, retry avec backoff
Secrets dans l'image DockerFuite si image partagéeSecret manager + injection runtime
Cold start Lambda/Cloud Run sans warmupLatence p99 catastrophiqueMin instances = 1, ou warmup ping
Pas de readiness probeTraffic vers pods non-prêts/ready vérifie TOUTES les dépendances
Scaling sur CPU uniquementQueue déborde, CPU stableAjouter métrique custom (queue depth)
Réponse synchrone > 30sTimeouts client, retries en cascadePasser en worker async + polling
Rate limit LLM API non géré429 en cascade, panne totaleRetry exponentiel + circuit breaker

Checklist de mise en production

  • Image Docker avec tag immutable, multi-stage, non-root
  • Secrets via secret manager (pas en dur, pas dans l'image)
  • /health et /ready implémentés et configurés dans Kubernetes/cloud
  • State 100% externalisé (Redis + DB)
  • Retry + circuit breaker sur les appels LLM API
  • HPA configuré avec métriques pertinentes
  • Métriques Prometheus exposées + dashboard Grafana
  • Rollback testé en staging avant chaque mise en prod
  • Canary ou blue/green pour les releases majeures
  • Alertes sur taux d'erreur > 1% et latence p95 > SLA

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.