We currently run agent sandboxes only via Docker (Lubuntu base image).
Cold starts are too slow and isolation still shares the host kernel. Firecracker microVMs give us VM-level isolation with container-like startup times and low memory overhead (boot in ≈125 ms, ≈5 MiB VMM overhead), which is a better fit for short‑lived AI agent sandboxes.
Goal: Introduce Firecracker as an alternative runtime backend, so each agent can choose between:
runtime = "docker" – today’s behavior.
runtime = "firecracker" – microVM per agent (or per session).
We currently run agent sandboxes only via Docker (Lubuntu base image).
Cold starts are too slow and isolation still shares the host kernel. Firecracker microVMs give us VM-level isolation with container-like startup times and low memory overhead (boot in ≈125 ms, ≈5 MiB VMM overhead), which is a better fit for short‑lived AI agent sandboxes.
Goal: Introduce Firecracker as an alternative runtime backend, so each agent can choose between:
runtime = "docker" – today’s behavior.
runtime = "firecracker" – microVM per agent (or per session).