A survey of 300 senior engineering leaders found that AI coding agents accelerate code production while increasing the burden of understanding, debugging and maintaining the result. Coleman Parkes conducted the research for debugging company Undo, so its findings should be read with the sponsor’s commercial interest in mind.
Respondents, most of whom oversee mission-critical C or C++ systems, reported spending an average of 9.8 hours per week producing code and 16.9 hours debugging development or production problems. Thirty-five percent of AI-generated code reached production before teams fully understood it, while 80 percent said agents struggled with difficult problems in complex codebases.
Seventy-nine percent agreed that agents generate code significantly faster. Yet they said the additional work required to inspect and repair that code meant the overall release cycle was no faster. Reports of hallucinated diagnoses, escaped defects and service incidents were also common, although the survey does not establish that AI caused every event.
The results are self-reported and concentrated in a demanding slice of software engineering, rather than a controlled productivity study. They still reinforce a practical point: generated code must remain reviewable, tested and understood, because typing speed is only one part of delivering reliable software.