Every agent that runs code needs a sandbox to execute it. The market now offers a dozen vendors with different billing models and performance claims that are hard to compare. Four properties decide the architecture and the real cost: cold start under concurrent load, filesystem persistence between turns, egress policy, and idle billing.
Cold start measured as time from create() to the first successful command shows Vercel Sandbox with the tightest burst latency (median ~0.7 s, P99 ~1.1 s). Modal and Runloop are close but have longer tails. Daytona can be sub‑300 ms when launched one‑by‑one but its success rate drops under burst, so retry logic is required. Cloudflare’s sandbox boots a full container and therefore shows a higher median (~5 s).
Filesystem persistence varies. E2B terminates the sandbox on timeout unless you set onTimeout:’pause’; Daytona keeps container sandboxes alive but clears memory on stop, while its VM sandboxes support pause/resume. Modal snapshots the filesystem with a seven‑day expiry. Cloudflare resets the disk after sleep unless you back up to R2. Vercel snapshots on stop, Runloop suspends state (Pro plan), Fly Sprites keep a 100 GB root disk, and Northflank offers persistent or ephemeral microVMs.
Egress policy can be denied everywhere, but the way conflicts are resolved differs. E2B favors allow rules, Vercel favors deny rules, and blocked TCP connections may appear open inside E2B; test with an application‑level response. Cloudflare, Vercel and Runloop let you change policy at runtime without restart.
Idle billing separates active‑CPU from wall‑clock charges. Vercel and Cloudflare bill only for used CPU, making them cheaper when the sandbox waits for a model. E2B, Daytona, Modal, Runloop and Fly Sprites bill wall‑clock unless you suspend or pause the instance. Suspend/resume on E2B, Fly Sprites and Runloop reduces idle cost dramatically when the gap between turns exceeds the pause overhead.
Pick the platform that matches your immovable constraint: waiting on models → Vercel Sandbox; kernel isolation or self‑host → E2B; persistence and forking → Daytona; GPU inside sandbox → Modal; Workers‑based egress security → Cloudflare Sandbox; coding agent with built‑in evals → Runloop; persistent per‑user machine → Fly Sprites; lowest rate plus BYOC → Northflank.
#AI #Product #Sandbox #DevTools #AgentEngineering #Cloud