Challenge environments for agent evaluation.

For agent researchers and AI engineers who want reproducible evaluation loops, not demo-grade scripts.