AWS re:Invent 2025 - Adding agentic AI to legacy apps with Amazon Bedrock AgentCore (MAM345)

Enhancing Legacy Applications with Agentic AI: AWS re:Invent 2025 Presentation Summary

Introduction to Agentic AI and Legacy Application Challenges

• Older, legacy applications are reliable but can be difficult to adapt or automate, especially for developers new to the codebase
• Agentic AI can interact with legacy systems and act/think on its own without rewriting the application

Understanding the Agentic Loop

• Agent receives a prompt, determines when/how to call an underlying language model (LLM)
• LLM generates responses based on training data
• Agent can also leverage external "tools" to perform tasks like web searches, database lookups, API calls, etc.
• This agentic loop of acting, interpreting, and planning continues until the task is completed

Leveraging Strands SDK and Amazon Bedrock AgentCore

• Strands is an open-source Python SDK for building AI agents with a "model-agnostic" approach
• Agents built with Strands can run on any infrastructure and call models hosted on Bedrock or OpenAI
• Bedrock AgentCore is a managed runtime and infrastructure offering from AWS to build, develop, and operate AI agents at scale

Integrating Agents with Legacy Applications via MCP

• MCP (Model-Centric Programming) is an open standard that defines how agents can securely connect and communicate with applications and services
• MCP exposes application capabilities as "tools" that agents can access, as well as "prompts" for data sources and "instructions" for the LLM

Demonstration: Building an Agentic AI Agent

1. Creating a Strands Agent: Importing the agent class, taking user input, and adding a simple calculator tool
2. Deploying to Bedrock AgentCore: Transforming the agent to run on the managed AgentCore runtime, including adding an HTTP server and entry point
3. Connecting to a Legacy Application:
   • Using AWS CodeGuru Developer to generate an OpenAPI spec for a legacy "Unicorn Store" application
   • Registering the OpenAPI spec as a "tool" in an AgentCore gateway
   • Updating the agent to fetch tools from the gateway and interact with the legacy app
4. Enhancing with Memory and Observability:
   • Leveraging AgentCore memory to maintain context across user sessions
   • Enabling tracing and observability to visualize the agent's interactions

Key Takeaways and Resources

• Agentic AI can be used to enhance legacy applications without rewriting the core system
• Strands SDK, Bedrock AgentCore, and MCP provide a comprehensive framework for building and deploying intelligent agents
• Observability and tracing are crucial for understanding and debugging agent behavior
• Further resources available in the provided GitHub repository with demos and tutorials