Agent Harness Lab Retrospective
An experiment on SWE-bench Lite comparing code repair by the same model using One-shot and an Agent Loop, with analysis of repeated rollouts, Pass@k, step limi…
Agent Harness Lab Follow-up Experiment Retrospective
A comparison of one-off generation, tool-free multi-turn revision, and an Agent that can search, edit, and run code on 20 real code-repair tasks from open-sour…
I Built a Video-to-Handout Tool with Codex. It Quietly Left Out Half a Lecture.
A look back at an AI Coding project and what I learned from it.