AWS re:Invent 2025 - Building AI Agents with Serverless, Strands, and MCP (NTA405)

Building AI Agents with Serverless, Strands, and MCP

Introduction to Agentic AI

• Limitations of standalone large language models (LLMs):
  • Passive, only aware of training data
  • Cannot take actions or access external systems
• AI agents overcome these limitations by:
  • Reasoning, tool discovery and selection, tool execution, and response generation
  • Enabling intelligent, autonomous behavior

Strands Agent SDK

• Simplifies building production-ready AI agents
• Uses state-of-the-art LLMs for reasoning, planning, and workflow design
• Supports any model that can execute tools, including Amazon Bedrock and Anthropic models

Key Components of an AI Agent

1. Model Access: Integrating the desired language model
2. Prompt: Defining the system prompt and agent persona
3. Tools: Accessing external tools and services

Integrating Tools with Strands

• Built-in tools like calculator, file system, and API integrations
• Custom tools defined using the @tool decorator
• Handling errors and fallbacks when external tools are unavailable

Model Context Protocol (MCP)

• Open standard for connecting AI applications to external tools and services
• Provides a common interface for AI agents to interact with diverse systems
• Strands has built-in support for MCP servers

Securing MCP Servers

• Importance of securing access to sensitive MCP servers
• Using AWS Lambda authorizers to validate API requests

Serverless AI Agent Architecture

• Deploying the AI agent on AWS Lambda, fronted by API Gateway
• Leveraging serverless benefits: scalability, security, observability

Demonstration Walkthrough

1. Weather Check: Using a built-in HTTP request tool to fetch current weather conditions
2. Commute Advisor: Implementing a custom tool to provide commute recommendations based on weather
3. Dice Roll: Integrating with an MCP server to roll dice and get the result

Key Takeaways

• Strands simplifies building production-ready AI agents with just a few lines of code
• MCP provides a standard interface for AI agents to interact with diverse external systems
• Securing MCP servers is crucial, and can be achieved using AWS Lambda authorizers
• Serverless deployment of AI agents offers scalability, security, and observability benefits

Business Impact and Applications

• Automating daily tasks and workflows with AI agents
• Integrating AI agents with enterprise systems and services
• Enhancing customer experiences with intelligent, context-aware assistants
• Rapid prototyping and deployment of AI-powered applications

Real-world Example

• The presented AI agent can:
  • Check the weather in Las Vegas
  • Provide commute recommendations based on weather conditions
  • Roll dice and return the result
• This demonstrates the agent's ability to intelligently orchestrate and execute multiple tools to fulfill a user's request.