I Cut AI Agent Handoff Payloads by 87%—Then the Important Questions Disappeared

Описание и применение

Сокращение объёма сообщений между ИИ‑агентами на 87 % за счёт разделения уведомлений и полного текста: короткие уведомления передаются в контекст, а полный текст хранится в «почтовом ящике» и извлекается по запросу. Приоритет вопросов гарантирует, что критические вопросы всегда доставляются в первом пакете.

Задача

Large AI agent handoffs were bloating context windows, causing repeated injection of full messages and missing critical unanswered questions in early turns.

Результат

Payload per handoff dropped from 969 bytes to 123 bytes (87.3 % reduction). The first notification batch now reliably contains at least one unanswered question, improving operational reliability.

Рабочий процесс

• 1. Hook receives full handoff text. 2. Hook generates a short notification containing type, state, sender, summary, and inbox path. 3. Notification is pushed into the active context. 4. Receiving agent pulls full message from inbox only when needed. 5. Prioritize unanswered questions, then state changes, answers, accepted work, progress/completion. 6. Enforce limits: max 5 notices, max 3 redeliveries, ~1,000 bytes budget, but always show highest‑priority item even if it exceeds budget.

Доказательства

Replaying 240 handoffs reduced the average payload per handoff from 969 bytes to 123 bytes: an 87.3% reduction.

Источник

After the change, the first batch carried at least one unanswered question instead of none.

Источник