EnvHarness Dynamically Adapts Environments for Agent Learning.

Large Language Model (LLM) agents learn through interaction with environments, but these environments are typically static and hand-built. This static nature means they often fail to adapt to an agent's evolving weaknesses and quickly become obsolete as the agent improves. While some environment generation methods exist, they often require domain-specific pipelines, rely on expensive verifiers, and still produce static outputs. To overcome these limitations and reduce the engineering burden, researchers propose EnvHarness. This is a programmable layer of plug-in components designed to wrap existing static environments, allowing their behavior to be reshaped dynamically without altering the original underlying logic. By operating through standard interfaces, EnvHarness is applicable across various domains while ensuring that every modified environment retains its original verification capabilities. Complementing EnvHarness is EnvRigger, an automated system that treats the target policy as a black box. EnvRigger observes agent execution trajectories to diagnose flaws and then synthesizes appropriate EnvHarness components to target these specific weaknesses. These synthesized components are then validated through fresh rollouts. Experiments across five benchmarks in four domains show that EnvHarness significantly outperforms both original environments and domain-specific generation pipelines, achieving up to a 9.0-point improvement on held-out instances with 9.8% fewer execution steps. It also provides a superior optimization signal for reinforcement learning, enabling continuous, targeted co-evolution of both the policy and its environment.