TalksAWS re:Invent 2025 - Build, deploy, and operate agentic architectures on AWS Serverless (CNS359)
AWS re:Invent 2025 - Build, deploy, and operate agentic architectures on AWS Serverless (CNS359)
Building Agentic Architectures on AWS Serverless
Overview
This session covered building, deploying, and operating agentic applications using AWS serverless technologies.
The presenters, Draj Mapatra and Hickey Park, are solutions architects at AWS.
The session assumed a 300-level understanding and focused on architectural principles and practical deployment patterns.
Customer Example: Retail Order Issues
The presenters used a customer example of issues with a retail order:
The customer was charged twice for the order
The customer only received one of the two ordered items
The customer needed the order quickly for a gift
This example was used to illustrate the challenges and potential solutions throughout the presentation.
Enhancing Customer Support with LLMs
The presenters discussed using large language models (LLMs) to provide better customer support and root cause analysis.
The initial approach involved using the LLM directly, providing it with a prompt and retrieving insights.
This was then enhanced by allowing the LLM to dynamically request additional context from various data sources (e.g., order history, inventory, payments) to improve the quality of the response.
Agentic Applications
The presenters introduced the concept of agentic applications, where the application logic is decomposed into autonomous agents that can dynamically reason about and execute tasks.
Agents have the ability to:
Invoke external tools and services to gather additional context
Make decisions about the next steps in the workflow based on the current state
Recursively execute this process until a final result is reached
Deploying Agentic Applications on AWS
The presenters discussed several patterns for deploying agentic applications on AWS serverless services:
AWS Lambda with API Gateway
AWS Fargate with Application Load Balancer
AWS Agent Core, a managed service for running agentic applications
Agentic Deployment Patterns
Monolithic Agent: A single, centralized agent that handles all tasks. This is a good starting point but can have scalability and maintainability challenges.
Serialized Pipeline: Agents are organized in a linear pipeline, with each agent responsible for a specific step. This can lead to bottlenecks if there are issues in one stage.
Orchestrator Agent: A central agent coordinates the execution of multiple specialized agents, each responsible for a specific domain or task.
Swarm Architecture: Multiple agents are deployed in parallel to perform broad research or exploration, then their results are synthesized. This can be more difficult to reason about and manage.
Security and Governance Considerations
Propagating security context and permissions across the agentic application is crucial.
Separate tokens should be used for the end-user, the agent, and any downstream API calls to maintain appropriate access control.
Governance policies and mechanisms should be applied consistently across the agentic application deployment.
Key Takeaways
Serverless technologies like AWS Lambda and Fargate can be effectively used to build and deploy agentic applications.
Agentic architectures allow for more dynamic, context-aware decision-making and task execution compared to traditional monolithic applications.
Decomposing agentic applications into specialized agents can improve scalability, maintainability, and flexibility, but requires careful architectural design.
Leveraging open standards like MCP (Model Context Protocol) and A2A (Agent-to-Agent) can simplify inter-agent communication and collaboration.
Integrating LLMs can enhance customer support and root cause analysis, but requires careful management of context and prompts.
Security, governance, and observability are critical considerations when deploying agentic applications in production.
These cookies are used to collect information about how you interact with this website and allow us to remember you. We use this information to improve and customize your browsing experience, as well as for analytics.
If you decline, your information won’t be tracked when you visit this website. A single cookie will be used in your browser to remember your preference.