Letter to Past Me: Three Mistakes That Cost a Day of Agent Runs

Описание и применение

Рецепт автоматизации, позволяющий избежать потери времени при работе с LLM‑агентом, используя изолированные ветки, журналирование времени и контроль за пустыми итерациями.

Задача

Repeated agent runs on a dirty working tree, unchecked token usage, and running loops on the same machine caused wasted time and lost work.

Результат

The recipe reduces wasted agent iterations from a full day to a few minutes, ensures reproducibility, and keeps sensitive data local.

Рабочий процесс

• Create a disposable worktree per prompt using fresh.sh
• Point the agent to the new worktree and capture the diff before cleanup
• Remove the worktree after each run
• Run ledger.sh to log wall‑clock time, exit status, and diff lines for each iteration
• Use gate.sh to stop the loop when an iteration produces zero diff lines
• Review the ledger to decide when to stop the loop and merge changes

Доказательства

The model was never the problem. My loop was.

Источник

The fix is a disposable worktree per prompt.

Источник

MonkeyCode is an open-source project that the operator describes as offering free model access and a free server option.

Источник