Bare IPython Harness Matches Production Coding Agents Using Qwen3.5-9B

An experiment testing small local coding agents demonstrated that a bare IPython harness matches heavy production harnesses on Terminal-Bench Lite, while tool-output normalization successfully eliminates most run-to-run variance. This finding proves that complex scaffolding is not always necessary for small local models, and highlights how environmental noise in tool outputs drastically impacts smaller language models. Researchers found that small models like Qwen3.5-9B are easily derailed by minor variations in command-line outputs, which can be fixed using a simple ~40-line regex normalizer and environment quieting.