Let's integrate AI Agents in Event-Sourced Systems — Divakar Kumar, FlyersSoft
Add agents beside deterministic fraud systems, not in place of them: project cross-service events into a semantic layer, send only uncertain cases to specialist agents, then emit the verdict back into the saga.
The design keeps rules and traditional ML as **tier one** for routine transactions, routing gray-zone cases to **tier two**. Events from transaction, account, device, and payment contexts feed a semantic projection used by risk and behavior agents; a third agent produces the verdict.
For an existing event-driven system, build agent context from read-optimized projections instead of querying raw event stores or collapsing bounded contexts. Return the result as another event so the agentic step participates in the established saga.
The design keeps rules and traditional ML as **tier one** for routine transactions, routing gray-zone cases to **tier two**. Events from transaction, account, device, and payment contexts feed a semantic projection used by risk and behavior agents; a third agent produces the verdict. For an existing event-driven system, build agent context from read-optimized projections instead of querying raw event stores or collapsing bounded contexts. Return the result as another event so the agentic step participates in the established saga. The demonstration used synthetic data and reported no fraud-accuracy, latency, or cost results. A third agent does not make a verdict deterministic, and the orchestration loop needs an explicit stopping condition to avoid running indefinitely.
This adapts agent orchestration to an existing event-sourced architecture instead of replacing it: deterministic rules handle routine cases, semantic projections provide bounded context for gray zones, and verdicts re-enter the saga as events. It reinforces staged automation and minimum-path routing, but the fixed multi-agent chain conflicts with evidence that specialist pipelines can lose coherence; synthetic demonstration and absent stopping rules leave reliability unproven.