I forked a live AI agent three ways, and every copy came up with its web server already running

Описание и применение

Эксперимент показал, что в Managed Agents можно чекпоинтить, форкнуть и откатить состояние микровМ, сохраняя запущенные процессы и файлы. Три форка стартовали с одним PID сервера, а откат восстановил убитый процесс с тем же PID.

Задача

How to preserve and explore an AI agent’s state efficiently in a disposable sandbox environment.

Результат

Forking a checkpointed session preserves all running processes and file state, allowing parallel exploration from the same exact state. Rollback restores killed processes with identical PIDs, confirming memory state is retained.

Рабочий процесс

• 1. Start a session and run a Python HTTP server inside the sandbox.
• 2. Build a single‑file web dashboard with an AI agent (OpenCode, DeepSeek v4 Pro).
• 3. Checkpoint the session (25.2 s).
• 4. Fork the checkpoint three times (15.3 s).
• 5. Verify each fork starts with the same PID 593 and serves the same index.html.
• 6. Give each fork a different design brief and observe the resulting dashboards.
• 7. Kill the server and delete the app directory in the parent session.
• 8. Roll back the parent session to the checkpoint (5.6 s).
• 9. Confirm the killed process restarts with PID 593 and serves the original page.

Доказательства

With the dashboard working, I checkpointed the session and forked it three times from that checkpoint:

Источник

fork1: PID 593 | local GET -> 200 | index.html 6695 bytes

Источник

rollback call: 5.59s until READY: 6.04s
app dir: index.html
server : PID 593
GET : 200

Источник