Developer forces AI agents into early live testing, surfacing 7 real Windows sandbox bugs

kevinkern · x · 2026-11-17

The author shares a workflow of forcing agents to do early live testing: agents produced a 57k-line test harness for complex tasks like building Windows sandboxes. His agent prompt: get the smallest real result working e2e first; stop after 6 agent hours without new learning and have a fresh reviewer cut scope; search the web for today's standard solution before big additions. The attached progress update shows 7 real product bugs found in the Windows agent sandbox, 6 already fixed on device.

Original post →

More from coding & agent

coding & agent channel →