Give every run its own machine — single-tenant, isolated, gone when it’s done. Run model-generated code without it touching your app or another user.
Every run gets its own VM — its own kernel, its own memory. One agent’s code never touches another user’s environment.
Save a clean baseline, run whatever the model generates, reset to clean for the next run.
Anything your agent builds and serves gets a URL you can open straight from the browser.
Boot from base, python, dev, or one you publish — no cold-start setup on every run.
Hold a session open across turns, or let it tear down clean. Pay only while it runs.
Connect agents directly via session-scoped MCP URLs — orchestration tokens and other sessions stay private.