NBER study revealed a gap between code generation and release frequency

A new study by NBER has shown that the introduction of AI tools in software development leads to a 240% increase in the number of commits, but the number of final releases increases by only 30%. The data indicates the emergence of a "bottleneck" in the software development life cycle (SDLC) processes, where the speed of writing code ceases to be the determining factor of success.

Analysis shows that developers spend significantly more time checking the results generated by agents than creating code from scratch. In particular, LinkedIn was forced to design a specialized platform for automated code review to cope with the increased volume of edits. Engineering teams are forced to rebuild workflows, shifting the focus from generation to verification and quality control.

For the effective integration of AI agents into production cycles, the creation of a comprehensive infrastructure is required. It includes strict specifications, management of the general project context, setting strict constraints, and implementing metrics that track the state of the code at each stage until release. Without these systems, automation leads to the accumulation of technical debt, rather than accelerating product delivery.

The current situation demonstrates the transition from the "AI assistants" stage to the "AI agents" stage, which requires deep integration into corporate CI/CD pipelines. Companies that do not adapt their verification and control processes face the fact that the increase in productivity at the code writing stage is offset by delays at the testing and integration stage.