AWS re:Invent 2025 - Architecting scalable and secure agentic AI with Bedrock AgentCore (AIM431)

Architecting Scalable and Secure Agentic AI with Bedrock AgentCore

Overview

• Presentation by Mark Brooker, VP and Distinguished Engineer at AWS
• Focuses on building reliable and secure AI agents using AWS Bedrock AgentCore

Inside an AI Agent

• AI agents are systems that take a goal, use an AI model for inference, and call external tools to achieve the goal
• Agents combine model inference with tool calls to handle tasks that models alone cannot do
• Agents can also include custom code to improve reliability, latency, and cost

Running Agents in Production

• Key components for running agents in production:
  • Agent Core Runtime: Secure, serverless environment to run agent code
  • Agent Core Gateway: Connect agents to external tools and services
  • Agent Core Memory: Persistent storage to remember user preferences
  • Agent Core Identity: Manage identities and credentials for agents
  • Agent Core Browser: Secure environment to automate web interactions

Agent Core Runtime

• Provides strong isolation using Firecracker microVM technology
• Scales serverlessly and charges based on active runtime, not idle time
• Allows any language, framework, or library to be used

Agent Core Memory

• Stores user preferences and conversation history to improve agent continuity
• Extracts relevant facts from conversations to provide to future interactions
• Helps agents remember important details about users

Agent Core Gateway

• Single place to connect agents to external tools and services
• Allows curation of tool sets to optimize agent performance
• Provides policy engine to control what agents are allowed to do

Neurosymbolic AI and Policy

• Combines neural and symbolic reasoning approaches for more powerful agents
• Agent Core Policy uses the CEDA policy language to formally specify allowed agent actions
• Provides mathematical guarantees about agent behavior and security

Evaluations

• Measures agent performance in production, not just in development
• Tracks metrics like goal success rate, conciseness, and tool call reliability
• Enables iterative improvement of agents based on real-world usage

Business Impact

• Enables developers to build reliable, secure, and scalable AI agents
• Unlocks new classes of AI-powered applications and services
• Provides the infrastructure to move agents from prototypes to production

Examples

• Personal assistant agent to automate daily calculations
• Outdoor activity planning agent to check weather, snow, and river conditions
• Secure isolation and policy enforcement to control agent behavior

Key Takeaways

• Agent Core provides the core infrastructure to run AI agents in production
• Strong isolation, memory, identity, and policy features enable reliable and secure agents
• Neurosymbolic AI and formal evaluations improve agent performance and safety
• Enables developers to rapidly build and deploy AI-powered applications