Researchers at Google introduce EnvHarness to address the problem that agent learning environments are static and don’t adapt to an agent’s specific weaknesses. The system provides a programmable plugin layer that dynamically reshapes existing environments without altering their underlying logic, paired with EnvRigger, which observes execution trajectories to automatically synthesize components targeting diagnosed flaws. Across five benchmarks, the approach achieved up to 9.0-point improvements on held-out test instances while cutting execution steps by 9.8%, enabling continuous co-evolution of policies and their training environments.
