Agentic Architecture Mediation for LLM Assistants: Preventing Solution-Jumping with Requirements Elicitation - Dataset

Generative AI assistants often exhibit solution-jumping: accepting a user's technology preference (e.g., ``I need Kubernetes'') without validating underlying requirements. For systems architects, this undermines critical engineering rigor. This article presents an Agentic Architecture Mediation System that steers LLM-based assistants toward requirements-first reasoning. By implementing a multi-agent Five Whys elicitation workflow and a deterministic evaluation engine, the system prevents premature technology commitment while preserving conversational accessibility. Testing across six industry anti-patterns demonstrated 100\\% detection of solution mismatches at an average cost of \\$0.01 per session. We show how to partition agentic responsibilities: using LLMs for dialogue and deterministic logic for evaluation to achieve robust, hallucination-free architectural guidance.