Off-Peak: the AI agent whose best feature is doing nothing

Описание и применение

Off‑Peak – это AI‑агент, который отложит обработку документов до тех пор, пока пользователь не будет отсутствовать, избегая прерываний и экономя энергию батареи.

Задача

Users want to process documents with an AI agent without interrupting their active laptop usage, conserving battery and avoiding interruptions.

Результат

Off‑Peak completed all 10 documents before the deadline, avoided 9 interruptions, and shifted 3.09 Wh of model energy to idle periods, demonstrating patient scheduling without user disruption.

Рабочий процесс

• 1. User provides a pile of documents and a deadline.
• 2. Off‑Peak monitors laptop idle time, battery, and thermal state via macOS/ioreg, pmset, psutil (or Linux equivalents).
• 3. It predicts work duration using a heuristic or TabPFN model.
• 4. If user is present and only_when_away is true, it waits until the latest safe start or until the deadline is reached.
• 5. When conditions are met, it processes up to two items per tick, writes output, checkpoints the item, and yields.
• 6. If the lid closes or the user returns early, it pauses after the current item and resumes from the last finished item.
• 7. After completion, it generates a report summarizing work done and energy usage.

Доказательства

Off‑Peak deferred 21 times, worked only in those away windows, and finished all ten before the simulated 10:30 PM deadline.

Источник

The baseline finished in 130 seconds and lost one document to a provider 503. Off‑Peak took 398 seconds of model time across the day.

Источник

The Gemma 4 31B endpoint returned intermittent 500/503s during the run. Three documents used Gemma; the rest fell back to Gemini 3.8 Flash.

Источник

The scheduler only sees a SystemState dataclass, so the same policy runs against real telemetry or the compressed day.

Источник