Google AI introduces EnvHarness to transform static agent environments into adaptive training systems

Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill introduced EnvHarness, an open-source framework turning static agent benchmarks into adaptive training environments. By wrapping environments in plug-in components, EnvHarness alters task conditions without modifying underlying simulators. Guided by the EnvRigger loop to write targeted wrappers, the system achieved up to a 9.0-point gain on ALFWorld and reduced SWE-bench Verified steps by 9.8%.

EnvHarness adaptive training framework

• ▪EnvHarness is released as open-source software under the Apache-2.0 license in Python, including reproduction drivers for six environments.
• ▪Researchers from Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill released EnvHarness, a programmable layer that turns static agent benchmarks into adaptive training environments.
• ▪EnvHarness requires a resettable environment to function, which rules out its application to live user accounts and physical robots.

Static environment limitations

• ▪Generating new environments to address static limitations is domain-specific and requires over-generating and heavily filtering untrustworthy LLM-written verifiers.
• ▪Traditional LLM agent training environments are hand-built and static, behaving identically regardless of which agent is acting or how much the agent has improved.

Component-based environment wrapping

• ▪EnvHarness wraps existing environments in plug-in components operating strictly through standard reset and step interfaces, leaving the underlying simulator, tasks, and human-built verifiers untouched.
• ▪EnvHarness includes three composable components: Stage to replay actions after reset, Contract to install per-step hooks, and Chain to compose a second environment into the same episode.

EnvRigger automated designer loop

• ▪EnvRigger rejects unsolvable and trivially solvable candidate wrappers, allowing up to five revision rounds per task, and compiles generated hooks in an isolated subprocess.
• ▪The EnvRigger LLM designer loop automatically writes EnvHarness wrappers by observing baseline rollouts, diagnosing systemic flaws, writing components in Python, and validating them on fresh rollouts.

Benchmark performance improvements

• ▪Across five benchmarks, including ALFWorld and SWE-bench Verified, skills mined using EnvHarness gained up to 9.0 points on held-out tasks with 9.8% fewer execution steps.
• ▪EnvHarness outperformed the domain-specific generator SWE-smith by 2.46 points on benchmarks while requiring 5.11 fewer execution steps.
• ▪On SWE-bench Verified, EnvHarness increased the resolved rate from 49.88% to 52.58% while reducing average execution steps from 55.01 to 49.61.

1 source