Bring an environment to exactly the state you want, then fan it out into parallel children in a single call. Collect the results, pay for the seconds you use.
Children start from the same live running point — not a fresh boot re-running setup.
Fan a session out into parallel children, each on its own isolated machine.
Pay for active compute only. Paused and idle sessions don’t run the clock.
Bring an environment to a known state once, then branch from it as many times as you need.
Sweeps, batch evals, RL rollouts — collect results across every child.
Python and TypeScript SDKs, the CLI, or remote MCP. No orchestration layer to run.