Replication Matters More Than Sharding in the AI Era

The practical workflow looks like this: an LLM agent proposes a database migration → the same architecture generates property-based tests for edge cases (concurrent writes, transaction rollbacks, duplicate Kafka events) → a separate step formally verifies the invariant before deployment, rather than after an incident.

At KT.Team, we build this kind of architecture with Python and Apache Kafka for the event layer, Elasticsearch for search and aggregations, and MCP as the protocol through which an AI agent securely accesses data—via an LLM & Security Gateway that logs and limits every call, rather than connecting directly to the production database. This is AI-native integration: an architecture with checks at every step between the agent and the data.

Appleton describes design engineering as a discipline at the intersection of design and development. A design engineer takes the interface through to a production build, bypassing the handoff of a design mockup to a separate developer. AI accelerates this intersection: generating components from design tokens and UI variations for A/B testing takes minutes instead of days. The discipline remains manual: a design-token system that does not sprawl across a project requires a human architectural decision, not a prompt.

The TTU principle—time to use—applies here. The value of an AI tool lies not in how impressive its demo looks, but in how long it takes to go from a request to a result you can trust under load. A seemingly simple outcome—stable replication, a correct zero-downtime migration, an interface without visual bugs—requires complex engineering under the hood: data models and verification processes that will not collapse under real traffic.

AI removed the constraint on coding speed. The constraint on the correctness of data systems remains. Companies that use this opportunity to hire prompt engineers without revisiting their replication architecture and invariant testing will get fast code and slow degradation under load. Companies that embed formal verification and event-driven architecture into the same AI architecture will get a system that holds up on the first high-traffic Friday—the kind of outcome business actually calls a result.