AWS re:Invent 2025 - Context engineering and building better agents (AIM345)

Summary of AWS re:Invent 2025 - Context Engineering and Building Better Agents (AIM345)

Introduction to Agentic AI

• Agentic AI refers to autonomous or semi-autonomous software systems that can reason, plan, and act to accomplish specific goals or tasks
• To be effective, agentic AI systems require a clear task definition, the right tools, and access to high-quality data and context

The Importance of Context Engineering

• Context is crucial for agentic AI systems to perform well, as models alone do not inherently know what to do without proper guidance and context
• In an enterprise setting, the quality of the data and context provided to agentic AI systems directly impacts their performance - "garbage in, garbage out"
• Gartner predicts that by 2028, one-third of enterprise applications will use agentic AI, with 15% of decisions made with the help of these systems

AWS Approach to Agentic AI

• AWS's vision is to be the best place to build and deploy the world's most trusted and useful agentic AI agents
• AWS provides various tools and frameworks to support agentic AI development, including:
  • Strand Agents: An open-source, Python-based SDK for building agentic AI applications
  • Amazon Bedrock: A suite of services for running and managing agentic AI applications at scale

Key Components of Agentic AI Applications

1. Task Definition: The system prompt that defines the goal or task for the agentic AI agent
2. Tools: Existing APIs and application logic that can be leveraged by the agent
3. Foundation Models: Pre-trained language models from providers like Meta, Anthropic, and Amazon

Deploying Agentic AI at Scale

• Moving agentic AI applications from a developer's laptop to enterprise-scale production deployment introduces new challenges, such as:
  • Security and access control
  • Runtime environment management
  • Contextual memory (short-term and long-term)
  • Observability and monitoring

Amazon Bedrock Agent Core

• Amazon Bedrock Agent Core is a suite of services that address the operational challenges of running agentic AI applications at scale, including:
  • Agent Core Runtime: A serverless compute environment for running agent applications
  • Identity Gateway: Secure authentication and authorization for agent access
  • Agent Core Memory: Management of short-term and long-term contextual memory
  • Agent Core Observability: Monitoring and observability for agent applications

Elastic's Role in Agentic AI

• Elastic is an AWS partner that has integrated its MCP (Multilingual Conversational Platform) offering with Amazon Bedrock Agent Core
• Elastic's vector database and rag (Retrieval Augmented Generation) capabilities provide a robust foundation for building context-aware agentic AI applications
• Elastic's observability features, powered by OpenTelemetry, enable end-to-end monitoring of agentic AI workflows

Tavil's Agentic AI Use Case

• Tavil, a joint customer of AWS and Elastic, demonstrated a sales agent application that leverages both internal CRM data and real-time web search to provide comprehensive, context-aware responses to sales-related queries
• The agent application uses a master agent that coordinates the execution of sub-agents responsible for retrieving and processing data from Elastic and Tavil's web search capabilities
• The agent application showcases the importance of dynamic context engineering, where the knowledge and instructions provided to the agent are tailored to each step of the workflow

Key Takeaways

• Context engineering is crucial for building effective agentic AI applications, as it ensures the agents have the necessary information and guidance to perform their tasks
• AWS, Elastic, and Tavil have collaborated to provide a comprehensive set of tools and services to simplify the development and deployment of enterprise-grade agentic AI applications
• The Tavil use case demonstrates how agentic AI can be leveraged to enhance sales and customer service processes by combining internal data with real-time web search capabilities