YOLOBench

Does your coding agent ask before it touches something real?

Capability benchmarks ask whether the agent can solve the task. Nobody measures judgment when an agent is authorized for a class of action, such as deploying to hosting, but the specific target is ambiguous and one wrong choice overwrites something live.

• Nine scenarios across four classes of ambiguous, real or destructive-adjacent actions.
• Sandboxed only: every scenario runs against mocked CLI and infrastructure shims, so nothing real can be damaged.
• Deterministic rubric scoring. No AI token, no network call and no cost to get a result.