/
Published on 02.04.2026
TLDR: Building an AI agent is the easy part — knowing if it actually works is where most engineers get stuck. This 7-part series walks through building real evaluation pipelines with datasets, evaluators, and a full harness integrated into your AI app.
TLDR: After shipping a vertical AI agent to production, the team discovered that many of their architectural choices — RAG, MCP layers, complex orchestration — were adding complexity without value. Stripping it all back made the product stable and fast.
The ZTRON Story: Why We Killed RAG and MCP
TLDR: When your technical documentation is deeply interconnected, standard RAG loses the relational context that makes that documentation useful. GraphRAG preserves those relationships, enabling agents to understand not just what documents say but how concepts within them relate to each other.