Google AI Introduces EnvHarness: A Programmable Layer That Turns Static Agent Environments Into Adaptive Training Worlds

EnvHarness is a programmable layer that modifies static training environments for AI agents by wrapping them with plug-in components that change where episodes start, what actions agents can take, and what they observe, while keeping the original human-built verification system intact. The tool matters because AI agents currently learn from hand-built environments that behave identically regardless of agent performance, limiting their ability to improve, and generating new environments requires domain-specific pipelines and unreliable automated verifiers. EnvRigger, an LLM component of the system, automatically diagnoses weaknesses in agent performance from rollout observation and writes targeted environment modifications, which are then validated on fresh rollouts. Across five benchmarks, skills trained in these adapted environments showed gains of up to 9.0 points on held-out tasks while requiring 9.8% fewer execution steps than skills trained in original environments.