Module 6: Deploy to AWS

Learning Objectives

• Understand Amazon Bedrock AgentCore Runtime
• Wrap a Strands agent for production deployment
• Deploy using the AgentCore starter toolkit
• Understand alternative deployment options (FastAPI + Docker)

Amazon Bedrock AgentCore

AgentCore is AWS’s managed platform for deploying AI agents. Key features:

Feature	Description
Runtime	Serverless hosting, up to 8-hour execution windows
Gateway	Convert APIs to MCP tools automatically
Identity	Auth via Cognito, Okta, Entra ID
Memory	Persistent cross-session memory
Observability	CloudWatch dashboards
Policy	Natural language guardrails
Evaluations	13 pre-built eval metrics

Option A: AgentCore Starter Toolkit

The fastest way to deploy.

1. Install the toolkit

   pip install bedrock-agentcore-starter-toolkit

2. Review the deployment entry point

   code module_06_deploy/app.py

   The key addition is the BedrockAgentCoreApp wrapper:

   from bedrock_agentcore.runtime import BedrockAgentCoreApp
   
   agent = Agent(
       system_prompt=SYSTEM_PROMPT,
       tools=[lookup_order, search_products, search_faq, create_support_ticket],
   )
   
   app = BedrockAgentCoreApp(agent=agent)

3. Configure the deployment

   cd workshop
   agentcore configure -e module_06_deploy/app.py

   This creates deployment configuration files targeting us-east-1.

4. Deploy to AgentCore

   agentcore launch

   Wait for deployment to complete (typically 2-5 minutes).

5. Check deployment status

   agentcore status

6. Invoke the deployed agent

   agentcore invoke '{"prompt": "What is your return policy?"}'
   agentcore invoke '{"prompt": "Check order ORD-10001"}'

Note

The AgentCore Runtime provides session isolation, so each invocation runs in its own environment. Sessions can last up to 8 hours for complex tasks.

Option B: FastAPI + Docker

For full control over the HTTP interface.

1. Review the FastAPI app

   code module_06_deploy/app_fastapi.py

   Key endpoints:

   • GET /ping: Health check (required by AgentCore)
   • POST /invocations: Invoke the agent

2. Run locally

   pip install fastapi uvicorn
   uvicorn module_06_deploy.app_fastapi:app --host 0.0.0.0 --port 8080

3. Test with curl

   curl -X POST http://localhost:8080/invocations \
     -H "Content-Type: application/json" \
     -d '{"prompt": "What is your return policy?"}'

4. Build Docker container (optional)

   docker build -t supportbot -f module_06_deploy/Dockerfile .
   docker run -p 8080:8080 supportbot

Deployment Architecture

flowchart TD
    subgraph AC[Amazon Bedrock AgentCore]
        subgraph RT[AgentCore Runtime]
            Agent[🤖 Your Strands Agent\nSupportBot]
            Agent --> Bedrock[Amazon Bedrock\nClaude Sonnet]
            Agent --> Mem[AgentCore Memory]
            Agent --> CW[CloudWatch Traces]
        end
        ID[🔐 Identity]
        GW[🔌 Gateway]
        POL[📋 Policy]
    end

    App[📱 Your Application\nWeb / Mobile / Slack] --> AC

Other Deployment Options

Target	How	Best For
AgentCore	agentcore launch	Managed, serverless
AWS Lambda	Wrap with Lambda handler	Short tasks, event-driven
AWS Fargate	Docker container on ECS	Long-running, custom infra
AWS EKS	Kubernetes deployment	Large-scale, existing K8s
Docker	Any container platform	Portability

CI/CD with GitHub Actions

For automated deployments, use the GitHub Actions workflow from AWS:

.github/workflows/deploy-agent.yml

name: Deploy SupportBot
on:
  push:
    branches: [main]
jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: aws-actions/configure-aws-credentials@v4
        with:
          role-to-assume: ${{ secrets.AWS_ROLE_ARN }}
          aws-region: us-east-1
      - run: |
          pip install bedrock-agentcore-starter-toolkit
          agentcore configure -e module_06_deploy/app.py
          agentcore launch

Key Takeaways

• AgentCore provides managed, serverless agent hosting
• The BedrockAgentCoreApp wrapper makes deployment straightforward
• For custom HTTP interfaces, use FastAPI + Docker
• AgentCore includes identity, memory, gateway, policy, and observability out of the box
• Use CI/CD for automated, production-grade deployments