Tag
A benchmark study on Terminal-Bench Lite finds that a bare IPython harness can match heavier production agent harnesses for Qwen3.5-9B, and that much of the run-to-run variance in small-model agents stems from noisy tool outputs rather than the model itself.
Nathan Cooper humorously explains how to use IPython as a terminal shell, highlighting features like running bash commands without the '!' prefix and displaying images in the terminal, with safety tools like safepyrun and safecmd.