What it is
Wraps static agent environments with modular interface-level components, then auto-customizes them from rollout failures to target policy weaknesses. Instead of generating new domain-specific environments, it wraps existing ones with plug-in state, interface, and task-composition transforms; a black-box rigger writes and validates these wrappers from agent rollouts while keeping the original verifier.
Key results
- Up to +9.0 points on held-out tasks vs original/generated envs
- 9.8% fewer execution steps while improving performance
- Up to +6.5 points in RL over learning in original environments
Targeted interface-level reshaping of existing environments can beat both static training worlds and domain-specific environment generation.