Reliable orchestration for small language models hinges on robust event flow and state management. The suggested architecture utilizes small model instances with minimal local state. Kafka serves as the event backbone, ensuring high-throughput delivery and partition-local ordering. Temporal acts as the orchestration layer, managing durable state and replaying execution after failures. Model invocations are treated as replayable side effects within this design. Temporal's Workflow Event History serves as the definitive record of conversational progress. While Kafka offers strong guarantees within its own transaction boundaries, external system interactions require different approaches. Correctness when interacting with external systems relies on idempotency, deduplication, sequence checks, and reconciliation. Global exactly-once semantics are not achievable across all components. This design prioritizes durability and replayability for stable orchestration.