Why coding agents have yet to deliver promised productivity gains

AI coding agents can generate code in minutes, but that speed has shifted work to later stages of software development. Engineers still need to understand the code, find bugs and investigate incidents, tasks that can become harder as codebases grow more complex.

The article argues that duplicated logic and layers of changes can make AI-generated code difficult to follow. It cites a survey in which nine in ten engineering leaders said code comprehension and debugging take longer in complex environments such as C/C++ codebases. Despite easier code generation, 79% said their release cycles were no faster.

Debugging agents may also lack evidence about what happened when a program ran. Source code, logs and documentation describe intended behavior, but may not show what occurred in a live system. The article says 93% of engineering organizations had experienced an issue whose root cause was incorrectly diagnosed because of an AI hallucination. It also says 37% of teams use AI agents for comprehension and debugging only with straightforward codebases.

The article points to a legacy software defect behind an outage at the UK’s national air traffic controller, which caused 2,000 flight cancellations and days of disruption. Its author, Undo co-founder and CTO Greg Law, argues that agents need better evidence about software behavior to help trace issues to their root causes. These figures and recommendations are presented in the article; no underlying survey details are provided in the supplied text.