The infrastructure agents reach for repeatedly — token counting, schema generation, data extraction — tends to get rebuilt from scratch each time, not out of carelessness but because writing 20 lines of custom code is faster than finding and integrating something external. You end up with dozens of slightly different implementations across a company's agent pipelines, each with their own edge case handling, none of them consistent with each other.
The author's argument about shared contracts applies here too. The agents that cause the most invisible damage aren't the ones doing novel things — they're the ones doing the same boring infrastructure work slightly differently, and the delta accumulates.