I run every AI agent run in its own Firecracker microVM. Here's what I learned. Sandscale, founded by a developer who found coding agents running unread commands on machines they cared about, now runs every AI agent execution in its own Firecracker microVM with a dedicated kernel and no shared state between runs. The founder notes that Firecracker does not filter network egress on its own, so network policy is handled as a separate layer, and billing is per awake vCPU-hour with paused and idle time between tool calls excluded. The service is in early access, priced at $0 for Free, $25/month for Pro and $48 per member per month for Team. Coding agents kept running commands I hadn't read, on machines I cared about. So each run in Sandscale now gets its own Firecracker microVM with its own kernel, and nothing is shared with the run next to it. If you're choosing between containers, gVisor and microVMs for code your agent wrote, I wrote up the tradeoffs here: https://sandscale.dev/blog/isolating-ai-agent-code-containers-gvisor-firecracker https://sandscale.dev/blog/isolating-ai-agent-code-containers-gvisor-firecracker . One thing worth knowing: Firecracker doesn't filter network egress by itself. Network policy is a separate layer. Billing is per awake vCPU-hour. Paused time and idle time between tool calls don't count against it storage above the allowance still bills . Free is $0, Pro is $25/month, Team is $48 per member per month. It's early access. I'd like to hear where it breaks for you. Try it: https://cloud.sandscale.dev/auth/signup https://cloud.sandscale.dev/auth/signup I'm the founder of Sandscale, so take my bias into account.