AWS re:Invent 2025 - Unlocking agentic AI access for microservices (ARC314)

Unlocking Agentic AI Access for Microservices

Introduction to Agentic AI and Microservices

• Agentic AI applications have emerged as a new pattern, with key differences from traditional microservices:
  • Highly stateful, data-intensive, and conversational
  • Require new DevOps and MLOps practices
  • Focus on enhancing human abilities through automation and insights
• Agentic applications can automate and streamline existing microservices and workflows
  • Example: Telco provider used an "S-Sur" agent to reduce incident response time from 30-40 minutes to 5 minutes

Microservices Principles and Challenges

• Key principles of microservices:
  1. Service independence
  2. Single responsibility
  3. Decentralized communication
  4. Agnostic tech stack
  5. System resiliency
  6. Independent deployability
• Challenges arise when integrating agentic AI with microservices:
  • Agents only understand "wide coding" and need context to interact with microservices
  • Exposing microservices logic in a structured, contextual way for agents to leverage

Infusing Agentic Capabilities into Microservices

• Agentic core constructs that can be infused into microservices:
  1. Reasoning: Agents' ability to solve problems and explain their chain of thought
  2. Memory: Short-term, long-term, and episodic memory to maintain context
  3. Learning: Agents' ability to improvise and improve through feedback loops
• Key differences from traditional, hardcoded microservices workflows:
  • Agents can dynamically determine which services to call and in what order to achieve a goal
  • Enables more flexible, adaptive, and self-improving automation

When to Use Agentic AI

• Start with real, high-impact business use cases (e.g. clinical workflow automation, market analysis)
• Define key metrics upfront (e.g. turnaround time reduction, staff productivity)
• Assess both business value and risk potential

Building Agentic Agents

• Agents are code (e.g. Python, TypeScript) that combine goals, instructions, and context
• Context engineering is crucial to guide agents' decisions on tools, data, and actions
• Agents need access to various resources:
  • Web scraping, navigation, and interactions
  • LLMs and text files
  • Workflows, automations, and knowledge bases
  • Structured and unstructured data

Model Context Protocol (MCP) for Agent-to-Tool Integration

• MCP is an open-source protocol for enabling seamless agent-to-tool communication
• Follows a client-host-server architecture:
  • Host: Agents, LLMs, and IDs requesting data access
  • MCP Client: Maintains 1-to-1 connection with MCP Server
  • MCP Server: Exposes microservices capabilities through standardized MCP protocol
  • Data Sources: Local and remote resources accessible via MCP
• Benefits of MCP:
  • Simplifies integration complexity compared to custom API integrations
  • Enables rapid development and deployment of AI features

Agent-to-Agent (A2A) Communication

• Agents need to discover, manage tasks, and collaborate with other agents
• A2A protocol addresses key challenges:
  1. Discovery: Mechanism to find relevant agents
  2. Task Management: Coordinating different agent capabilities
  3. Collaboration: Sharing context, security, and other metadata

Exposing Microservices for Agentic Access

• Define and describe APIs (e.g. using Swagger, OpenAPI, gRPC)
• Implement security and guardrails for accessing microservices
• Incorporate feedback loops, observability, and retry mechanisms

AWS Agent Core Platform

• Key components of the AWS Agent Core platform:
  1. Agent Core Runtime: Serverless environment for running agents
  2. Agent Core Memory: Stores short-term, long-term, and episodic context
  3. Agent Core Identity: Provides secure, delegated access to enterprise services
  4. Agent Core Observability: Comprehensive visibility into agent execution and performance
  5. Agent Core Gateway: Exposes existing APIs, Lambda functions, and MCP servers to agents
• Enables rapid development and production-ready deployment of agentic applications

Strands: Open-Source SDK for Building Agents

• Strands is an open-source SDK that simplifies agent development:
  • Ease of use: Just 4 lines of code to create a working agent
  • Robust capabilities: Pre-built tools and MCP/A2A support
  • Extensibility: Supports multiple LLM providers

Key Challenges and Considerations

• Cardinality and token costs: Managing interactions with thousands of microservices
• Auditability requirements: Handling policy changes and precise, deterministic logic
• Integration with existing DevOps and MLOps practices

Conclusion

• Business-driven use cases are key, focusing on high-impact problems and defined metrics
• Integrating agentic AI with microservices requires careful planning and best practices
• AWS Agent Core platform provides a modular, production-ready solution to unlock agentic access