Skip to content

Lesson 14: Deploying Agents to the Cloud

Play

Watch on YouTube

Code for this lesson: samples/14-deploy

From Local to Production

Section titled “From Local to Production”

Everything so far has run on a local machine. A Strands agent is just Python, so you can containerize it and run it anywhere. This lesson deploys it to AWS using Amazon Bedrock AgentCore: Runtime for isolated execution, Memory for persistent and long-term recall, and Observability for automatic tracing.

AgentCore Overview

Making Your Agent Deployable

Section titled “Making Your Agent Deployable”

Wrap the agent in a BedrockAgentCoreApp and give it an entrypoint. The agent itself doesn’t change; the customer service agent below is the same one built across the previous lessons, with an AgentCore Memory session manager swapped in for persistence:

from strands import Agent
from bedrock_agentcore.runtime import BedrockAgentCoreApp
from bedrock_agentcore.memory.integrations.strands.config import (
    AgentCoreMemoryConfig, RetrievalConfig
)
from bedrock_agentcore.memory.integrations.strands.session_manager import (
    AgentCoreMemorySessionManager
)


app = BedrockAgentCoreApp()


MEMORY_ID = os.environ.get("BEDROCK_AGENTCORE_MEMORY_ID", "")


def create_agent(actor_id: str, session_id: str):
    # AgentCore Memory for persistence + long-term recall
    session_manager = None
    if MEMORY_ID:
        config = AgentCoreMemoryConfig(
            memory_id=MEMORY_ID,
            session_id=session_id,
            actor_id=actor_id,
            retrieval_config={
                "/users/{actorId}/facts": RetrievalConfig(),
                "/users/{actorId}/preferences": RetrievalConfig(),
            },
        )
        session_manager = AgentCoreMemorySessionManager(
            agentcore_memory_config=config,
            region_name="us-east-1",
        )


    return Agent(
        tools=[lookup_customer, get_order_history, process_refund],
        plugins=[AgentSkills(skills=["./skills"]), RefundWorkflowHandler(), tone_handler],
        system_prompt=SYSTEM_PROMPT,
        context_manager="auto",
        session_manager=session_manager,
    )

📂 main.py · lambda-deployment/ (alternative: AWS Lambda)

The retrieval_config namespaces tell AgentCore Memory which long-term memory strategies to pull from on each turn, so the agent recalls facts and preferences about a returning user without replaying old conversations.

Deployment Commands

Section titled “Deployment Commands”
Terminal window
# Install CLI
npm install -g @aws/agentcore


# Scaffold project
agentcore create


# Add memory
agentcore add memory --name CustomerServiceMemory --strategies SEMANTIC,USER_PREFERENCE


# Deploy
agentcore deploy


# Invoke
SESSION_ID="<SESSION ID HERE>"
agentcore invoke --session-id "$SESSION_ID" \
  '{"prompt": "I need help returning my order", "actor_id": "sarah"}'

What AgentCore Provides

Section titled “What AgentCore Provides”
ComponentWhat It Does
RuntimeIsolated microVMs, maintains state across requests
MemoryConversation persistence + semantic long-term recall
ObservabilityAutomatic traces, logs, execution telemetry
GatewayExpose APIs as MCP-compatible tools
IdentityIAM or OAuth authentication gate

AgentCore has more capabilities beyond these. See the AgentCore developer guide for the full picture.

What You Still Need

Section titled “What You Still Need”

The deployed agent is only part of a production architecture. You still need API gateways, rate limiting, retries, security controls, monitoring, and error handling around it. The surrounding architecture matters as much as the agent.

Resources

Section titled “Resources”