AWS re:Invent 2025 - Building AI Agents with Kiro, MCP, and Amazon Bedrock AgentCore (DEV331)

Building AI Agents with Kiro, MCP, and Amazon Bedrock AgentCore

Overview

• This presentation covered the tools and techniques for building AI agents, including Kiro, the Agentic IDE, the Machine Call Protocol (MCP), and Amazon Bedrock AgentCore.
• The speakers, Eric Henchin and Dwan Lightfoot, are developer advocates at AWS focused on agentic AI.
• The goal was to provide the audience with the knowledge and tools to start building their own AI agents today.

What are AI Agents?

• AI agents are autonomous systems that leverage AI to reason, plan, and take actions to achieve goals on behalf of humans.
• Agents need access to APIs, tools, memory (short-term and long-term), and the ability to maintain context and state across conversations.
• Deploying agents securely with the right access controls and observability is a key challenge, which is where Amazon Bedrock AgentCore comes in.

Amazon Bedrock AgentCore

• AgentCore is a serverless runtime for deploying and managing AI agents, providing:
  • Secure, isolated runtime for agent workloads
  • Short-term and long-term memory management
  • Gateways for integrating existing APIs and tools
  • Identity and access management for agents
  • Observability and monitoring through Amazon CloudWatch
• AgentCore supports any AI model or framework, making it flexible for building diverse agent applications.

Kiro - The Agentic IDE

• Kiro is a new IDE from AWS focused on building agentic applications.
• Key features include:
  • Spec-driven development to define agent requirements and have Kiro generate the implementation
  • Integrated support for MCP servers, including the ability to add the AgentCore MCP server
  • Vibe mode for rapid, iterative development
  • Agent hooks to automate actions based on IDE events

MCP (Machine Call Protocol)

• MCP is a standard for exposing AI-callable APIs and tools.
• MCP servers can provide access to a wide range of functionality, from documentation lookup to custom business logic.
• AgentCore includes a dedicated MCP server to integrate with, providing agents access to a curated set of tools.
• Semantic search in AgentCore gateways can help limit agent access to only the tools it needs, reducing token usage and cost.

Demo Application

• The presenters walked through a demo application that included:
  • A "Food Tracker" agent as the main user-facing interface
  • A "Recipe" agent that orchestrated calls to two MCP servers:
    1. A recipe server with a database of recipes
    2. A nutrition/ingredients server
  • Leveraging AgentCore memory management to maintain context across sessions
  • Observability features in AgentCore to monitor agent performance and behavior

Key Takeaways

• Building effective AI agents requires integrating multiple components - AI models, APIs, memory, and observability.
• AgentCore provides a comprehensive platform to simplify the deployment and management of agents, addressing the "undifferentiated heavy lifting".
• Kiro IDE streamlines agent development with features like spec-driven design and MCP integration.
• Observability is crucial for understanding agent behavior, performance, and cost, especially as agents become more complex.
• The demo showcased a real-world example of how these AWS tools can be used to build a multi-agent system for a practical application.