MCP Recipe to Spot Over-Engineered Code.

How It Works

Your AI agent reads the GitHub repository: a distributed event processing pipeline using Kafka, Avro schemas, and a custom dead-letter queue with exponential backoff.

The agent runs `validate_radical_simplification`. Jargon Elimination: 'event-driven choreography with CQRS' becomes 'services send messages to each other instead of calling each other directly, and we read data from a different place than we write it.' Core Mechanism: the ONE thing making this work is message ordering guarantees , without partition-key ordering, the entire pipeline produces incorrect results.

First Principles: the pipeline exists because writes happen 100x more frequently than reads, so separating them reduces contention. Self-Deception: the team claims 'exactly-once processing' but the dead-letter queue retry mechanism can produce duplicates when the consumer crashes between processing and acknowledgment.

Complexity Justification: Kafka is justified (high throughput needed), but the custom DLQ is not , a standard retry topic with TTL achieves the same result with zero custom code.

The agent posts to #engineering in Slack: a 3-paragraph brief that any engineer , junior or senior , can understand, with the verdict highlighting the self-deception about exactly-once semantics.