EnvHarness: Dynamic Reconstruction of Static Environments

Researchers from Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill proposed EnvHarness, which wraps static benchmark environments for LLM agent training without modifying internal code, transforming them into training environments targeting specific weaknesses.

Limitations of Existing Environment Generation

Existing automated environment generation pipelines are domain-specific, making them non-transferable across fields such as web, programming, and tool use. Additionally, ensuring the accuracy of validators generated by LLMs is difficult, and logic errors in generated tools can hinder training.

How EnvHarness Works

EnvHarness is a programmable layer that applies the concept of an Agent Harness to the environment side. It preserves the original environment's validators while transforming only the state, action, and observation spaces.

• Stage: Manipulates initial states to insert obstacles requiring specific skills or to adjust task difficulty.
• Contract: Redefines the action space, transition dynamics, and observation space to change interaction rules.

Performance Improvements

In experiments across 4 domains and 5 benchmarks, agents trained with EnvHarness showed up to 9.0 points improvement on unseen tasks compared to training on original environments, with interaction steps decreasing by 9.8%. It also generates stronger policies in reinforcement learning, demonstrating sustained performance gains in both human-created and generated environments.