A work item lands on a queue. It might be a document to classify, a case to route, a knowledge update to inspect, or an exception that needs review. At first, it is tempting to treat that as a simple model call: send the payload to an LLM, get a result, and move on. At production scale, that decision is rarely so simple.

This session focuses on inference architecture for event-driven agents: when inference should run, what context it needs, what result event it should emit, and how the workflow should behave when the model is uncertain or wrong. We will walk through a practical pattern: deterministic pre-checks validate the event, inference runs only when model use is justified, results emit new events with confidence and trace metadata, low-confidence outcomes route to human review, and failures land in dead-letter flows that become operational and governance signals.

From there, we will connect the pieces into a production architecture: queues for discrete inference work, streams and topics for multi-consumer state changes, agent memory or vector indexes as derived context, and warehouses or lakehouses as the feedback layer for evaluation, drift review, audit, and cost tracking. The patterns apply across streaming-backed architectures where events carry context, trigger inference, and preserve a record of what happened.

The core question is not whether agents should use events. It is how events should control inference: what should trigger a model call, what should stay deterministic, what should be persisted or replayed, what should become evaluation data, and where a human needs to stay in the loop. Attendees will leave with a practical framework for designing event-driven inference workflows that are observable, reliable, and governable in production.