Sandswap: Ultra-Efficient Sandboxes for AI Agents

🤖 Sandswap: Ultra-Efficient Sandboxes for AI Agents
📅 Tuesday, November 10th, 2026
🕖 07:00 CET / 11:30 IST / 14:00 Beijing & SGT / 17:00 AEDT
Presenter: Chaoyi Huang, original author of Sandswap

Sandswap rethinks the microVM: different AI agents can share and reuse a single isolated microVM in milliseconds. Ideally a microVM can be reused 183 times per second by different AI agents, and an agent can be rotated through a microVM in as little as ~6 ms.

In Sandswap, strongly isolated compute environments (microVMs running only a kernel and a pid 1 warden) are decoupled from data storage (SandFS, the root filesystem of an AI agent). A SandFS is switched into the microVM dynamically for command execution, with latency as low as ~2.3 ms cold and ~0.96 ms warm. Since an agent's thinking time is on the order of seconds, every bit of microVM idle time can be reused by other agents. That is how Sandswap reaches very high resource reuse and cuts the cost of agent execution environments by tens to hundreds of times.

The session covers the underlying mechanism and the tough parts: keeping latency overhead ultra-low in a large-scale system, and maintaining VM-grade isolation and security while SandFS is dynamically mounted into and unmounted from microVMs.