AI Agent Development Is Severely Imbalanced: Focused on Programming Tasks, Ignoring Most Labor Markets

A large-scale study shows that current AI agent development focuses almost exclusively on programming tasks, ignoring most labor markets. Researchers from Carnegie Mellon University and Stanford University compared 43 agent benchmarks with the U.S. labor market and found that programming-related fields, which account for only 7.6% of total U.S. employment, dominate agent development. The study reveals that highly digitized fields like management (88% digitization rate), law (70%), and architecture/engineering (71%) have extremely low representation in benchmarks (1.4%, 0.3%, and 0.7% respectively). Low-wage fields such as personal services and care are also barely covered. In terms of skills, agents only focus on information acquisition and computer operations, covering less than 5% of the job market, with almost no coverage of interpersonal interaction skills. Researchers point out that this imbalance stems from methodological convenience. They call for new benchmarks to cover fields like management and law, design more realistic and complex tasks, and promote refined evaluation. OpenAI's GDPval benchmark is cited as a positive example for its wide coverage of fields.