EnvHarness: Awakening Static Worlds for Agent Learning

TL;DR - EnvHarness is a programmable wrapper that dynamically reshapes static agent-learning environments without altering their underlying logic or verifiers. Its automated EnvRigger system diagnoses policy weaknesses from execution trajectories and generates targeted components, improving both evaluation performance and reinforcement-learning signals.

• Uses plug-in components and standard interfaces to adapt environments across multiple domains while preserving original verifiers.
• Treats the target policy as a black box, synthesizing modifications from observed trajectories and validating them with fresh rollouts.
• Across five benchmarks in four domains, achieved up to a 9.0-point improvement on held-out instances while using 9.8% fewer execution steps.
• Supports continuous co-evolution of an agent policy and its training environment through targeted reinforcement-learning feedback.