Undo Survey Finds AI Coding Agents Do Not Speed Up Releases

A survey of 300 senior engineering leaders managing mission-critical software, conducted by Coleman Parkes on behalf of software debugging firm Undo, reveals that AI coding assistants are failing to accelerate the overall software delivery pipeline. Although 79 percent of respondents acknowledged that AI agents generate code significantly faster, the time saved is lost to debugging and comprehension. Consequently, engineering leaders report that the entire software release cycle remains no faster than it was before the introduction of AI.

The survey, which focused primarily on teams working with C and C++ codebases, highlights a stark imbalance in how developers spend their time. On average, engineering teams spend 9.8 hours per week writing code, but they dedicate 16.9 hours per week to debugging issues, which consumes 42 percent of their typical work week. This shift has introduced severe risks to production environments. According to the report, 35 percent of AI-generated code is deployed to production before developers fully understand how it works.

This comprehension gap has led to widespread technical issues. The survey found that 81 percent of organizations experienced at least one production incident or service outage in the past six months, with 14 percent suffering these disruptions multiple times a month. Furthermore, 93 percent of respondents reported at least one instance of incorrect troubleshooting due to AI hallucinations, and 18 percent faced this issue multiple times per month. Additionally, 91 percent experienced test escapes or poorly optimized code entering production, with 8 percent seeing this happen multiple times a month.

For software practitioners, the findings suggest that AI agents are currently ill-equipped for complex systems, as 80 percent of leaders noted that agents struggle with difficult problems in complex codebases. About one-third of teams restrict AI agent usage to straightforward codebases. Undo chief executive officer Greg Law noted that developers frequently spend days trying to untangle code that is "almost, but not quite right." Ultimately, the study underscores that software engineering remains less about raw code generation and more about understanding system architecture and managing trade-offs.