Building Agents at Cloud Scale — Antje Barth, AWS

Key Concepts

• AI Agents at Cloud Scale
• Generative AI Applications
• Alexa Reimagining
• Agentic Services
• Amazon Q Developer
• Strands Agents (Open-Source Python SDK)
• Model-Driven Approach
• Amazon Bedrock Integration
• Multi-Agent Workflows
• Memory and Retrieval Augmented Generation (RAG)
• Multi-Modal Support (Images, Video, Audio)
• Machine Communication Protocol (MCP)
• AWS Lambda Integration
• Streamable HTTP
• Agent-to-Agent Communication (A2A)
• Useful General Intelligence (UGI)

Amazon's AI Transformation and Alexa Reimagining

Amazon believes AI will reinvent customer experiences. They have over 1,000 generative AI applications in development, impacting inventory forecasting, delivery route optimization, and customer shopping experiences. A key example is the complete reimagining of Alexa, representing a large integration of services, agentic capabilities, and LLMs.

• Alexa Plus: Works through hundreds of specialized expert systems (groups of capabilities, APIs, and instructions) and orchestrates across tens of thousands of partner services and devices.
• Scale: Over 600 million Alexa devices are currently in use.

Building Agentic Services: Amazon Q Developer

Amazon Q Developer is a code assistant that helps across the software development lifecycle. A developer agent for the CLI was released, bringing an agentic chat experience to the terminal. It helps debug issues, answer natural language questions, and read/write files.

• Development Time: The Amazon Q Developer agent for CLI was built and shipped in just three weeks.

Strands Agents: A Model-Driven Approach

To enable rapid development of AI agents, Amazon developed a model-driven approach, encapsulated in the Strands Agents open-source Python SDK. This approach leverages the reasoning, planning, and action-taking capabilities of LLMs, allowing developers to focus on the agent's purpose rather than the implementation details.

• Strands Agents: Connects the model and the tools.
• Integration: Integrates with Amazon Bedrock by default (using Cloud 3.7 Sonnet), but also supports other providers like Llama, Anthropic, Meta, and OpenAI through LightLLM integration. Custom model providers are also supported.
• Tools: Comes with over 20 pre-built tools for file manipulation, API calls, AWS service integration, memory, and RAG.
• Retrieve Tool: Enables semantic search over a knowledge base. An internal AWS agent uses this to manage over 6,000 tools by storing their descriptions in a knowledge base.
• Multi-Modality: Supports images, video, and audio.
• Multi-Agent Workflows: Supports graph-based workflows and swarms of sub-agents.
• MCP Integration: Natively integrates with MCP, allowing connection to thousands of MCP servers.

Machine Communication Protocol (MCP) and AWS Lambda Integration

MCP is crucial for agent-to-agent communication. AWS is actively involved in the MCP community, contributing code and helping to evolve the protocol.

• Standard IO: Initially, MCP servers used standard IO for local connections.
• Streamable HTTP: To enable remote connections and scaling, AWS uses streamable HTTP and deploys MCP servers as Lambda functions behind an API Gateway.
• Security: Integrates with authorizers (e.g., Cognito) for security and authorization. Session data can be stored in DynamoDB.
• MCP Lambda Handler: A dedicated MCP Lambda handler simplifies setting up MCP servers in Lambda.
• Example: A demo showed a "roll dice" tool hosted as a Lambda function, accessible through Strands Agents via MCP.

Agent-to-Agent Communication (A2A) and the Future

The next step is enabling agents to communicate with each other. AWS is excited about emerging open protocols like MCP.

• Agent Stores: The vision is a future where personal agents connect to agent stores, enabling agents to work together to accomplish tasks.
• Atomic Unit of Digital Interactions: Danielle stated that "The atomic unit of all digital interactions will be an agent call."

Conclusion

The presentation highlights Amazon's commitment to AI agents at scale, showcasing examples like the reimagined Alexa and the Amazon Q Developer agent. Strands Agents, an open-source Python SDK, empowers developers to rapidly build production-ready AI agents using a model-driven approach. The integration with MCP and AWS Lambda enables scalable and secure agent-to-agent communication, paving the way for a future where agents seamlessly collaborate to fulfill user needs.