AWS re:Invent 2025 - A practitioner’s guide to data for agentic AI (DAT315)

Transitioning to Agentic AI: A Practitioner's Guide to Data and Governance

The Evolving Landscape of AI Agents

• Rapid advancements in AI over the past few years, from simple chatbots to more advanced agents with context, hybrid search, and autonomous capabilities
• Example of buying car insurance used to illustrate the progression from 2023 to 2025

The Three Pillars of Agentic AI

1. Reasoning: Using large language models (LLMs) to understand and process user requests
2. Action: Executing tasks and calling tools to fulfill user needs
3. Memory: Maintaining context and history to enable iterative, conversational interactions

The React-Loop: Reason, Act, and Manage Context

• Agents go through a loop of reasoning, planning actions, and managing context
• Maintaining and compacting context is crucial to enable long-running, stateful interactions

Agentic Memory Management

• Short-term memory (agent state), medium-term memory (semantic and episodic), and long-term memory (prompts)
• Leveraging caching at different levels to optimize performance and reduce costs

Introducing MCP: Model Context Protocol

• Standardizing the interaction between agents and the tools they call
• Agents can discover and invoke tools through the MCP server

Choosing the Right Tools for the Job

• Taxonomy of tools: general-purpose vs. specialized, read-only vs. data mutation
• Aligning tool selection with user personas (data analysts, developers, end-users)

Connecting Agents to Data Sources

• Accessing data from various databases, data warehouses, and streaming sources
• Leveraging the MCP server to interact with different data stores

Addressing Data Challenges with a Data Marketplace Architecture

• Identifying data producers and consumers within the organization
• Building data products and exposing them as APIs to be consumed by agents

Governing Agentic AI Experiences

• Ensuring data quality, discoverability, fine-grained access control, and performance
• Propagating trusted identities and entitlements through the agent ecosystem

A Reference Architecture for Agentic AI

• Leveraging AWS Bedrock Agent Core as the runtime for agent applications
• Integrating LLMs, knowledge bases, and specialized data processing engines
• Exposing data sources as MCP tools for secure, governed access

Call to Action

• Focus on building specialized data APIs to provide secure, governed access
• Prioritize end-to-end data governance and security from the start
• Leverage AWS services and tools to simplify the agentic AI experience