Same Agents, Different Minds — What 180 Configurations Proved About AI Environment Design

Google tested 180 agent configurations. Same models. Same tasks. Error rates differed by 3.9x. The environment isn't plumbing — it's the model.