33 broken builds, one agent on each disposable computer. The hard part was deciding what counts as fixed

Описание и применение

В ходе хакатона 33 проекта не собрались. Автоматический агент исправил 17 из них, изменив в среднем 4 строки кода, при этом машина была заморожена с подробным отчётом для команды.

Задача

Hackathon submissions often fail to build end‑to‑end; teams receive only a score, not actionable debugging information.

Результат

The agent fixed 17 of 33 broken builds with an average of 4 source lines changed, costing ~$7 in model time per machine and ~$16 total.

Рабочий процесс

• 1. Rewind frozen machine to failure point
• 2. Install coding agent on machine
• 3. Provide agent with failure command, last 40 output lines, and instructions
• 4. Agent attempts minimal change to make command exit 0
• 5. Harness reruns command to verify success
• 6. If successful, agent writes FIX.md and machine is frozen again

Доказательства

17 of 33 now build.

Источник

The median fix was 4 lines.

Источник

All 33 machines cost about $7 in model time.

Источник