AWS re:Invent 2025 - Service-oriented builders guide to agentic AI: Insights from WEX (ARC313)

Transitioning Software Builders to Agentic AI: Insights from WEX

Embracing Agents in Service-Oriented Architectures

• Agents are not a radical departure from traditional service-oriented architectures (SOA)
• Agents leverage similar design principles like loose coupling, modularity, and reusability
• Agents introduce new concepts like persistent memory, autonomous coordination, and reasoning transparency
• Builders can leverage their existing skills in distributed systems to succeed with agentic AI

Key Design Principles for Agentic Systems

1. Loose Coupling: Agents should be deployed and scaled independently, with asynchronous processing.
2. Modularity: Break down agents into granular, specialized components with clear responsibilities.
3. Reusability: Build MCP servers and tooling to be reused across multiple agents and use cases.
4. Discoverability: Maintain catalogs and documentation for agents, their capabilities, and integration points.
5. Stateful Memory: Embrace persistent contextual memory for multi-turn conversations and workflows.
6. Orchestration vs. Autonomous Coordination: Balance deterministic workflows with emergent coordination at runtime.
7. Service Contracts and Capability Emergence: Allow agent capabilities to evolve without breaking client integrations.
8. Goal-Oriented Design: Focus on achieving specific goals rather than just CRUD operations.
9. Reasoning Transparency: Ingest and process the evidence of an agent's reasoning process for observability.
10. Self-Correction: Understand and manage an agent's tendency to self-correct and find alternative paths to achieve its goals.
11. Non-Determinism: Account for the inherent non-determinism in agent behavior when testing and operating the system.

WEX's Journey with Agentic AI

• WEX, a global commerce platform, used Bedrock and Agent Core to build agentic systems for operational support
• Started with a chatbot "Chat GTS" to automate repetitive support requests and provide self-service capabilities
• Expanded to event-driven agents that can autonomously triage, investigate, and remediate infrastructure issues
• Agents leverage MCP servers, Kendra knowledge bases, and Bedrock action groups to interact with existing tools and services

Key Lessons Learned

1. Architecture Still Matters: The fundamentals of distributed systems design still apply when building agentic AI.
2. Start Small, Think Big: Begin with simple use cases, but architect for scalability and extensibility.
3. Maintain Work-Life Balance: It's easy to get caught up in the rapid pace of AI innovation, so remember to take breaks and maintain perspective.

Technical Details and Results

• Utilized AWS services like API Gateway, Step Functions, DynamoDB, and S3 to build the agentic AI platform
• Integrated with Google Chat for user interaction and Active Directory for access control
• Achieved over 2,000 internal users for the agentic support system within the first year
• Reduced operational overhead and improved incident response times by automating repetitive tasks

Business Impact

• Freed up human support engineers to focus on more complex, high-value work
• Improved operational reliability and resilience by proactively detecting and remediating issues
• Increased self-service capabilities, allowing users to find information and resolve problems independently
• Established a foundation for continued expansion of agentic AI across the organization

Example Use Cases

1. Network Troubleshooting: An agent can autonomously investigate network issues, diagnose the problem, and provide a detailed report to the user, reducing the need for manual intervention.
2. Infrastructure Incident Response: An agent can detect infrastructure anomalies, open support tickets, execute remediation runbooks, and update stakeholders, all without human involvement.