A bare IPython harness ties production coding agents with Qwen3.5-9B on TB-Lite, and tool-output normalization fixed most of the run-to-run variance

We’ve been testing which agent harness works best for small local models. The result: a single-tool IPython harness can match much heavier production harnesses on Terminal-Bench Lite, and much of the “randomness” in small-model agents comes from noise in the environment, not the model. Setup

中文全文翻译正在补齐，可先查看下方英文原文。

We’ve been testing which agent harness works best for small local models. The result: a single-tool IPython harness can match much heavier production harnesses on Terminal-Bench Lite, and much of the “randomness” in small-model agents comes from noise in the environment, not the model. Setup