Undo survey says C/C++ teams spend 42 percent of the week debugging

An engineering team that got twice as fast at writing code may still ship on the same cadence as last year, because the time it saved went straight into finding out why the new code does not work. Undo, a debugging tool vendor, commissioned the independent research firm Coleman Parkes to survey 300 senior engineering leaders, with fieldwork done in July and August 2026, and InfoQ has now summarised the results. Ninety-three percent of the sample work with C/C++, and every organisation has more than $250 million in revenue, so this is a mission-critical codebase picture rather than an all-developers picture.

The debugging number, with the method

Across a 40-hour week, respondents say they spend 9.8 hours on producing code, 16.9 hours on debugging issues identified during development, and 7.5 hours on debugging production issues, with 6.8 hours on other tasks. InfoQ reports debugging during development as "42% of the average working week", the figure for the 16.9-hour slice.

Code comprehension is the other large piece. Thirty-five percent of AI-generated code reaches production before the team has fully understood what it does, per the report. Ninety-three percent of leaders say coding agents struggle with difficult problems in complex codebases, and Undo quotes 91 percent saying code comprehension and debugging take longer in complex C/C++ codebases. Nine of ten teams had at least one production incident affecting customers in the previous six months, with 14 percent reporting multiple incidents per month.

The claim that is doing most of the work

The striking claim is at the end. Undo reports that 79 percent of leaders say "the resulting shift in effort toward debugging and 'unpicking' AI-generated code means the overall release cycle is 'no faster than before'", so the time saved on code generation is reappearing somewhere else and releases happen at the same cadence as before.

Two caveats sit on this. The vendor sells AI-powered root-cause analysis, which is the product that becomes necessary when the finding is true. Coleman Parkes ran the survey independently but the question set was commissioned by Undo. And the sample is 300 senior leaders in large C/C++ shops: the finding is strongest for mission-critical codebases where "code must be understood", per Undo's wording, and weaker for a greenfield TypeScript team shipping once a week.

The failures the leaders admit

Three failure modes carry their own numbers in the report. Incorrect root-cause diagnosis caused by model hallucination hit 93 percent of teams at least once in six months, with 18 percent reporting it multiple times per month. Test escapes or poorly optimised code reaching production hit 91 percent at least once, with 8 percent reporting it multiple times per month. About one-third of teams "use AI agents for comprehension and debugging only in straightforward codebases", per InfoQ, falling back on manual methods in complex ones.

Greg Law, founder and CEO of Undo, writes that engineers "lose days trying to unravel what went wrong and why" with "code that's almost, but not quite right", and that "while agents are great at writing reams of code quickly, they're less capable at debugging it". That is the vendor's own framing of the finding, read it as such.

The useful part for a reader is the breakdown of the hours, which is checkable against your own team's week. If generation at your shop is 9 hours and debugging is 10, the lever shifts. If debugging is 20, the finding fits.

Source

AI Coding Agents in Complex Codebases: Survey Finds Increased Debugging and a Comprehension Gap, Sergio De Simone, InfoQ. Primary data: Overcoming the Limitations of Coding Agents in Complex Software Systems, Undo, with research conducted by Coleman Parkes.