Deploy to production

Getting an agent into production means choosing where it runs and wrapping it in the entry point that host expects. The agent you built in development is the one you ship: the same loop, tools, and model provider run unchanged behind whatever target you pick here. What changes between targets is the packaging around it, not the agent itself.

Deployment targets

Amazon Bedrock AgentCoreRun on the managed AgentCore runtime for agents.

AWS LambdaServerless, event-driven invocations with no servers to manage.

AWS FargateRun containers on ECS without provisioning hosts.

AWS App RunnerShip a container as a scaling web service from source or image.

Amazon EKSRun on managed Kubernetes when you need cluster control.

Amazon EC2Run on instances you manage for full control of the host.

DockerContainerize the agent as the foundation for any host.

KubernetesDeploy the container to any Kubernetes cluster.

TerraformProvision the agent's infrastructure as code.

Nx Plugin for AWSScaffold and deploy an agent project with the Nx tooling.

A deploy-ready agent

Most targets call the agent over HTTP: they send a request to an invocation route and check a health route to know the container is live. The agent behind those routes is the same one you build anywhere else.

Python

from fastapi import FastAPI
from strands import Agent

app = FastAPI()
agent = Agent()


@app.post("/invocations")
async def invoke(request: dict):
    result = agent(request["prompt"])
    return {"output": result.message}


@app.get("/ping")
def ping():
    return {"status": "healthy"}

TypeScript

import { Agent } from '@strands-agents/sdk'
import express, { type Request, type Response } from 'express'

const agent = new Agent()
const app = express()
app.use(express.json())

app.post('/invocations', async (req: Request, res: Response) => {
  const result = await agent.invoke(req.body.prompt)
  res.json({ output: result.lastMessage })
})

app.get('/ping', (_: Request, res: Response) => res.json({ status: 'healthy' }))

app.listen(Number(process.env.PORT) || 8080)

Each target guide takes this shape and adds the packaging it needs: a container image, a Lambda handler, a Kubernetes manifest, or a Terraform module.

Where to go next

New to deployment? Start with Docker to containerize the agent locally, then move the image to a cloud target. Already know where you are shipping? Pick that target from the grid above and follow its guide end to end.

Before you ship, read Operating agents in production for the practices that apply across every target: configuration, secrets, scaling, and observability.

Tags

deployment15