A team from Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill has released EnvHarness, a programmable layer that transforms static agent benchmarks into adaptive training worlds. The system wraps existing environments using plug-in components that operate strictly through the standard reset() and step() interface, allowing the environment to adapt to the policy's training without altering the underlying simulator or human-built verifiers.

  • EnvHarness uses an LLM designer called EnvRigger to diagnose flaws in a policy's rollouts and automatically write targeted wrappers.
  • The system includes three composable components: Stage (rewrites start states), Contract (installs per-step hooks on actions and observations), and Chain (composes second environments under shared budgets).
  • Across five benchmarks including ALFWorld, WebArena, SWE-bench Verified, OfficeQA, and SpreadsheetBench, skills mined via EnvHarness gain up to 9.0 points on held-out tasks.
  • On SWE-bench Verified, the approach reduces average execution steps by 9.8% while increasing resolution rates from 49.88 to 52.58.
  • The code is available as Apache-2.0 Python, requiring only a resettable environment and an existing agent evaluation loop.

The authors consider this important because static environments stop teaching once solved, whereas EnvHarness allows agents to continue learning by targeting specific weaknesses identified during training.