Making smaller Minecraft agents measurable

Craftwright treats traces, verified skills and repeatable scenarios as the proving ground instead of relying on a single impressive demo.

A Minecraft agent can look convincing while still failing to understand whether an action actually worked. Craftwright makes observation, execution and verification separate parts of the trace.

Skills return structured outcomes and benchmark scenarios provide the same initial conditions across model and runtime changes. That makes regressions visible instead of anecdotal.

The benchmark is not the destination. It is how the agent runtime earns enough trust to become useful outside a staged demonstration.