No matter how an "AI agent" is defined (and improved upon), the paper argues that complex reasoning will "collapse" - more research suggests AI tool usefulness could be measured with a half-life.