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
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.
Proof1 community thread
Key points
- We’ve been testing which agent harness works best for small local models.
- Harnesses: QwenCode, OpenHands, opencode, Pi, plus a bare ipython session.
- QwenCode just routes those reads through dedicated tools because its prompt explicitly says to use readfile/grepsearch/globinstead of cat/grep/find.
- v0: The IPython kernel was separate from the task environment, which confused the model.
Sources (1)
- [1]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 variancer/LocalLLaMA (top, daily) · Oct 19, 12:40 PM
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.
We’ve been testing which agent harness works best for small local models.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026Qwen Flash Next quant and temp comparison - performance vs task time vs quality
- Oct 17, 2026deepseek-ai/DeepSeek-V4.1-Flash-1017
- Oct 16, 2026nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-MXFP8
- Oct 16, 2026Qwen Flash Next potential SOTA Q2 nobody talks about
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026Character swap with Minimax in Comfy - Cat Fiction
