← back
Let's integrate AI Agents in Event-Sourced Systems — Divakar Kumar, FlyersSoft
Takeaway
Integrate agents into existing event-sourced workflows by giving ambiguous cases coordinated access to relevant cross-domain evidence.
Summary
- FlyersSoft augments existing rule-based and ML fraud detection with a second-tier agent for ambiguous transactions, preserving the systems that handle straightforward cases.
- The agent needs real-time evidence spread across transaction, account, device, and payment bounded contexts, including merchant details, compliance status, fingerprints, and chargebacks.
- An orchestration layer coordinates these contexts using asynchronous messaging and hosts the agentic decision process.
- Cosmos DB stores an append-only history of domain and integration events, with change-data capture providing a route for downstream processing.
event-sourcingfraud-detectionagent-orchestration
Original description
A card gets declined and no one, including the customer, can say exactly why. That gray zone is where Divakar Kumar points his agents. In a payments and fraud system, a rule based engine and an ML model already score most transactions cleanly; the hard cases are the ambiguous ones that neither can resolve. His approach adds an agentic layer on top of an existing event sourced architecture rather than replacing it, so the bounded contexts already in the system, transaction, device, and account, become the context the agents reason over. Events flow through change feeds into projections and a semantic layer that the agents read, communicating asynchronously through a message broker in a saga style loop. A risk analyzer agent, a second agent that reaches a verdict, and a third work the case while guarding against infinite loops and keeping memory short, all runnable serverless. The takeaway is architectural: event sourcing already carries the state and history an agent needs, so the cleanest way to add judgment to a production system is to layer agents onto the events you are already emitting. Speaker info: / divakar-kumar https://iamdivakarkumar.com Timestamps: 0:00 - Introduction: adding agents to an existing system 1:20 - A declined transaction you can't explain 3:04 - Where rule based and ML systems fall short 5:40 - Handling the gray zone with agents 5:53 - Bounded contexts: transaction, device, account 8:24 - Event sourcing and change feeds 10:57 - Building the semantic layer 13:16 - Avoiding infinite loops 14:07 - The risk analyzer and verdict agents 15:24 - The saga orchestration loop 19:00 - Putting the architecture together