AWS re:Invent 2025 - Cox Automotive's Blueprint for Agentic AI on Amazon Bedrock AgentCore (IND3329)

Cox Automotive's Blueprint for Agentic AI on Amazon Bedrock AgentCore

Introduction

• Presentation by Ravi (AWS Solutions Architect), Brian Lloyd Newubber (Cox Automotive), and Tabari Goen (Cox Automotive Lead Architect)
• Focuses on Cox Automotive's journey in adopting agentic AI and their blueprint for building successful agentic applications

Challenges of Deploying AI Agents

• Key challenges in moving AI agents from prototype to production:
  • Scalable and cost-effective runtime infrastructure
  • Maintaining context and conversational state
  • Secure access and integration with existing systems
  • Visibility and observability into agent behavior
• Example use case: Automotive service technician assistant agent

Amazon Bedrock AgentCore

• Fully managed services for building and deploying agentic AI applications:
  • Runtime: Secure, scalable, and serverless execution environment
  • Memory: Short-term and long-term memory management
  • Identity: Secure authentication and credential management
  • Gateway: Integration with existing APIs and Lambda functions
  • Observability: Visibility into agent execution and behavior

Cox Automotive's Agentic AI Journey

• Overview of Cox Automotive's business and data assets
• Adoption of generative AI in 2023, leading to several successful products
• Shift to agentic AI in 2024, with initial challenges in driving adoption

Accelerating Agentic AI Adoption

• "Start with crazy and work backwards" - Cox Automotive's approach
• Key principles:
  1. Set a clear goal and deadline to drive excitement and momentum
  2. Leverage existing assets (AWS partnership, Bedrock, in-house expertise)
  3. Focus on a single agentic framework (Strands on Amazon Bedrock AgentCore)
  4. Emphasize simplicity in agent development (writing prompts, configuring tools)

Building a Production-Ready Agentic Solution

• Case study: Cox Automotive's automated virtual assistant for customer conversations
  • Orchestrator routing messages to domain-specific sub-agents
  • Leveraging Bedrock AgentCore for data isolation and framework flexibility
• Key patterns and considerations:
  1. Start with a solid foundation (Bedrock AgentCore)
  2. Implement comprehensive red teaming and safety checks
  3. Use hard and soft guardrails to balance safety and user experience
  4. Automate evaluation of agent performance and behavior
  5. Implement circuit breakers to prevent runaway costs and behaviors

Lessons Learned and Recommendations

1. Get moving - don't wait for the "perfect" solution, start experimenting
2. Think disruptively - use agentic AI to fundamentally change assumptions and capabilities
3. Red team early and often to understand system failures and vulnerabilities
4. Design for the worst-case scenario and build evaluation frameworks accordingly
5. Implement hard limits and circuit breakers to ensure graceful failure

Resources

• QR codes for:
  • Leveling up agentic AI skills
  • Bedrock AgentCore deep dive hands-on workshops