A practical guide to deploying AI agents to production. Covers Docker, serverless, platform deployments, plus essential production concerns: health checks, monitoring, error handling, rate limiting, and cost control.

All deployment methods share a common FastAPI application. Build this first, then deploy anywhere.

agent-service/
├── main.py              # FastAPI app
├── requirements.txt
├── Dockerfile
├── docker-compose.yml
└── fly.toml             # (optional) Fly.io config

import hashlib
import os
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from openai import AsyncOpenAI
from tenacity import retry, stop_after_attempt, wait_exponential
 
app = FastAPI(title="Agent Service")
client = AsyncOpenAI(api_key=os.environ["OPENAI_API_KEY"])
 
# --- Models ---
class AgentRequest(BaseModel):
    prompt: str
    max_tokens: int = 1000
    user_id: str = "anonymous"
 
class AgentResponse(BaseModel):
    result: str
    tokens_used: int
 
# --- Retry logic for LLM calls ---
@retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=2, max=10))
async def call_llm(prompt: str, max_tokens: int) -> dict:
    response = await client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
        max_tokens=max_tokens
    )
    return {
        "content": response.choices[0].message.content,
        "tokens": response.usage.total_tokens
    }
 
# --- Health checks ---
@app.get("/health")
async def health():
    return {"status": "healthy"}
 
@app.get("/ready")
async def ready():
    try:
        # Verify LLM API is reachable
        await client.models.list()
        return {"status": "ready"}
    except Exception as e:
        raise HTTPException(status_code=503, detail=f"Not ready: {e}")
 
# --- Main endpoint ---
@app.post("/invoke", response_model=AgentResponse)
async def invoke_agent(request: AgentRequest):
    try:
        result = await call_llm(request.prompt, request.max_tokens)
        return AgentResponse(
            result=result["content"],
            tokens_used=result["tokens"]
        )
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Agent error: {str(e)}")

fastapi>=0.115.0
uvicorn[standard]>=0.32.0
openai>=1.50.0
pydantic>=2.9.0
tenacity>=8.2.0
slowapi>=0.1.9
prometheus-fastapi-instrumentator>=6.1.0
redis>=5.0.0

FROM python:3.12-slim
 
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
 
COPY . .
EXPOSE 8000
 
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]

services:
  agent:
    build: .
    ports:
      - "8000:8000"
    environment:
      - OPENAI_API_KEY=${OPENAI_API_KEY}
    depends_on:
      - redis
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3

  redis:
    image: redis:7-alpine
    ports:
      - "6379:6379"
    volumes:
      - redis_data:/data

  prometheus:
    image: prom/prometheus
    ports:
      - "9090:9090"
    volumes:
      - ./prometheus.yml:/etc/prometheus/prometheus.yml

  grafana:
    image: grafana/grafana
    ports:
      - "3000:3000"
    environment:
      - GF_SECURITY_ADMIN_PASSWORD=admin

volumes:
  redis_data:

# Build and run
docker compose up --build -d
 
# Test
curl http://localhost:8000/health
curl -X POST http://localhost:8000/invoke \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello, agent!", "user_id": "test"}'

Fly.io deploys containers globally with edge compute and auto-scaling.

app = "my-agent-service"
primary_region = "iad"
 
[build]
  dockerfile = "Dockerfile"
 
[http_service]
  internal_port = 8000
  force_https = true
  auto_stop_machines = true
  auto_start_machines = true
  min_machines_running = 1
 
  [http_service.checks]
    [http_service.checks.health]
      interval = "30s"
      timeout = "5s"
      path = "/health"
      method = "GET"
 
[env]
  # Non-secret env vars here
 
[[vm]]
  size = "shared-cpu-1x"
  memory = "512mb"

# Install flyctl, login, then:
fly launch --no-deploy
fly secrets set OPENAI_API_KEY="your-key"
fly deploy
 
# Scale
fly scale count 3
fly scale vm shared-cpu-2x --memory 1024

Modal is ideal for AI workloads — scales to zero, supports GPUs, Python-native.

import modal
 
app = modal.App("agent-service")
image = modal.Image.debian_slim().pip_install(
    "fastapi", "uvicorn", "openai", "pydantic"
)
 
@app.function(
    image=image,
    secrets=[modal.Secret.from_name("openai-secret")],
    container_idle_timeout=300,
    allow_concurrent_inputs=10
)
@modal.asgi_app()
def serve():
    from main import app as fastapi_app
    return fastapi_app

# Deploy
modal deploy modal_app.py
 
# Logs
modal app logs agent-service

Cost-effective for low-traffic agents. Use Mangum to adapt FastAPI to Lambda.

# lambda_handler.py
from mangum import Mangum
from main import app
 
handler = Mangum(app)

pip install mangum
# Deploy via SAM, CDK, or Serverless Framework
# Set OPENAI_API_KEY in Lambda environment variables

Lambda limits to be aware of: 15 min timeout, 10 GB RAM max, cold starts of 1-3s.

Railway auto-deploys from GitHub with built-in databases.

# Procfile
web: uvicorn main:app --host 0.0.0.0 --port $PORT

1. Connect your GitHub repo to Railway
2. Add environment variables (OPENAI_API_KEY)
3. Railway auto-detects Dockerfile or Procfile and deploys
4. Get a *.railway.app URL instantly

# Add to main.py
from prometheus_fastapi_instrumentator import Instrumentator
 
instrumentator = Instrumentator(
    should_group_status_codes=True,
    should_respect_env_var=False
)
instrumentator.instrument(app).expose(app, endpoint="/metrics")

global:
  scrape_interval: 15s

scrape_configs:
  - job_name: "agent-service"
    static_configs:
      - targets: ["agent:8000"]

# Add to main.py
from slowapi import Limiter, _rate_limit_exceeded_handler
from slowapi.util import get_remote_address
from slowapi.errors import RateLimitExceeded
from starlette.requests import Request
 
limiter = Limiter(key_func=get_remote_address)
app.state.limiter = limiter
app.add_exception_handler(RateLimitExceeded, _rate_limit_exceeded_handler)
 
@app.post("/invoke", response_model=AgentResponse)
@limiter.limit("10/minute")
async def invoke_agent(request: Request, agent_request: AgentRequest):
    # ... same implementation
    pass

import redis.asyncio as redis
 
redis_client = redis.from_url("redis://redis:6379")
 
async def check_budget(user_id: str, tokens_requested: int, daily_limit: int = 100_000):
    key = f"tokens:{user_id}:{datetime.now().strftime('%Y-%m-%d')}"
    current = int(await redis_client.get(key) or 0)
    if current + tokens_requested > daily_limit:
        raise HTTPException(
            status_code=429,
            detail=f"Daily token budget exceeded ({current}/{daily_limit})"
        )
 
async def record_usage(user_id: str, tokens_used: int):
    key = f"tokens:{user_id}:{datetime.now().strftime('%Y-%m-%d')}"
    await redis_client.incrby(key, tokens_used)
    await redis_client.expire(key, 86400)  # Expire after 24h

import hashlib
 
async def get_cached_response(prompt: str) -> str | None:
    prompt_hash = hashlib.sha256(prompt.encode()).hexdigest()
    cached = await redis_client.get(f"cache:{prompt_hash}")
    return cached.decode() if cached else None
 
async def cache_response(prompt: str, response: str, ttl: int = 3600):
    prompt_hash = hashlib.sha256(prompt.encode()).hexdigest()
    await redis_client.setex(f"cache:{prompt_hash}", ttl, response)

• ☐ Health check endpoints (/health, /ready)
• ☐ Structured logging (JSON format)
• ☐ Retry logic with exponential backoff
• ☐ Rate limiting per user/IP
• ☐ Token budget enforcement
• ☐ Response caching for repeated queries
• ☐ Prometheus metrics + Grafana dashboards
• ☐ HTTPS enforcement
• ☐ Input validation (Pydantic models)
• ☐ Error handling that never leaks internals
• ☐ Secrets in environment variables, never in code

• How to Build a RAG Pipeline
• How to Evaluate an Agent
• How to Use MCP

deployment docker fly-io modal serverless fastapi production how-to