Each task provides an isolated environment, instructions, and a verifier that checks the resulting state. Agents must use terminal tools across multiple steps rather than return only a textual answer. Reproducible containers, time limits, and tool permissions are part of the evaluation contract.