One Bug, Five Engineering Layers: A Practical Guide to LLM Agent Design

A deliberately broken three-line Python duration parser — which misreads 250ms as 250.0 seconds instead of 0.25 — serves as the basis for comparing five distinct LLM engineering approaches. The five layers are prompt, context, harness, loop, and graph engineering, each differing in what is designed and how the system responds to failure. The framework draws on an August 2026 preprint titled 'Graph Engineering in the Era of LLM Agents,' which proposes graph engineering as a method for coordinating tasks, agents, and runtime state. Tools like LangGraph support this model by allowing workflow graph nodes to contain either model calls or conventional code. The article uses a fixed input-output contract for the parser to objectively evaluate repairs produced at each engineering layer.