Google researchers have open-sourced EnvHarness, a framework that changes an AI agent’s training environment in response to its recurring weaknesses. Instead of rewriting the model, benchmark or verifier, it wraps an existing environment and adjusts what the agent sees and which actions are allowed.
The framework plugs into the standard reset and step interface used by many reinforcement-learning environments. Its setup components alter an initial state, rule components constrain transitions after an action, and link components change the observations returned to the agent. A companion system called EnvRigger reviews behavior, identifies failure patterns, proposes targeted modifications and validates them through new interactions.
The goal is to avoid a fixed benchmark becoming uninformative once an agent learns its repeated tasks. Google says the approach improved results on WebArena, SWE-bench Verified and ALFWorld for models including Gemini 3.1 Flash-Lite, Qwen3.6 27B, Gemini 3.5 Flash and Claude Sonnet 4.6. Reported gains ranged from 2.9 to 3.7 percentage points.
Those figures are research results, not evidence that the framework improves every production agent. Dynamically changing an environment can also complicate comparisons between runs. EnvHarness’s practical contribution is a reusable layer for generating harder, weakness-specific practice while keeping the original evaluator intact.