Why Are Coding Agents So Dumb?

Coding agents, despite improvements in underlying AI models, remain significantly limited in their practical capabilities. The author identifies key shortcomings including poor task management (executing parallel tasks sequentially instead of simultaneously) and inability to delegate work efficiently, making agents the bottleneck in AI-assisted development rather than the models themselves. While the distinction between models (the "brain" generating code) and agents (the "body" executing it) is crucial, current agent software has failed to advance at the pace of the models powering them.