Why AI agents need your existing event store, not a new architecture

Divakar Kumar presents a real-world architecture for integrating AI agents into existing event-sourced systems, specifically for fraud detection. The key insight is that rule-based and ML models fail in the 'gray zone' of transaction risk, where agents can add context from multiple bounded contexts.

The approach uses a tiered system: Tier 1 handles clear cases with existing rules/ML, while Tier 2 uses agentic AI for uncertain transactions. Agents access a semantic layer built from domain events across transaction, account, device, and payment contexts via CDC or message brokers.

The architecture employs a fan-out pattern with specialized agents (risk analyzer, behavior analyzer) and a verdict agent that synthesizes their outputs. This avoids the false positives common in traditional systems by incorporating broader contextual data through tool-based access to materialized views.

Practical considerations include strict latency requirements (sub-500ms), using short-term memory only, and implementing loop termination metrics. The POC demonstration shows events propagating through CDC to projection layers, with agents accessing this semantic layer to produce decisions that feed back into the saga orchestration.