ANALYSIS Two developments this week quantify a structural problem in AI-assisted software development: speed gains at the point of code generation are being consumed by debugging, comprehension, and compliance work downstream. A 300-person survey of mission-critical engineering leaders and a new governance platform built for code moving at "agent speed" put numbers and product bets behind that thesis.
Why it matters
The enterprise pitch for AI coding tools rests on a productivity multiplier. But productivity measured at the point of generation ignores the rest of the delivery pipeline. 79% of engineering leaders said AI agents can generate code significantly faster2. Yet those leaders said the overall release cycle is no faster than before because effort shifts toward debugging and unpicking AI-generated code3. ◆ When the vast majority of leaders report faster generation but no faster releases, the speed gains are being consumed elsewhere in the process. Separately, SCANOSS launched Earnie, a governance platform designed to catch open-source license violations introduced by agents that have "no visibility into a team's license policy, and no reason to pause and check one"4.
The big picture
The Coleman Parkes survey covered 300 senior engineer leaders responsible for delivering mission-critical software, most of whom working with C/C++, across the UK and US. The headline finding: 35% of AI-generated code reaches production before development teams fully understand what the code does. Debugging accounts for 42% of the average working week in those environments, with teams spending an average of 16.9 hours per week debugging compared with 9.8 hours per week producing code. As Undo founder and CEO Greg Law put it, agents are "great at writing reams of code quickly but less capable at debugging it," and engineers are "being buried in an avalanche of code that is beyond human capacity to debug".
The failure data is stark. 91% of surveyed teams experienced test escapes, serious defects, or poorly optimized code entering production at least once. 93% experienced AI hallucinations leading to incorrect diagnoses of code problems. And 81% experienced a production incident or service outage affecting internal users or customers at least once in the previous six months. 55% of respondents said agents introduce incorrect code too frequently, creating rework and delivery delays.
On the compliance side, SCANOSS CEO Julian Coccia framed the problem in blunt terms: "An agent commits as fast as it can generate tokens, with no idea what your policy is. Earnie is what makes sure obligations don't slip through just because review didn't have time to catch up". Earnie checks license obligations down to the copied snippet at pre-commit and at merge, organized around dated releases rather than point-in-time scans7. Each customer works in an isolated environment, with only fingerprints and hashes shared with SCANOSS for matching6.
ANALYSIS These two stories describe the same structural problem from different angles. The Coleman Parkes survey quantifies the debugging and comprehension gap; SCANOSS addresses the legal-compliance gap. Both exist because AI agents optimize for token generation without downstream awareness, whether that downstream is runtime behavior or license obligation. The survey's finding that 80% of respondents said coding agents struggle to solve difficult problems in complex codebases reinforces the pattern: agents are most useful where the work is least consequential, and least reliable where the stakes are highest.
The commercial interests of the messengers deserve noting. Undo sells debugging tools, and SCANOSS sells software governance, so both benefit from framing agent-generated code as a problem requiring their solutions. The survey numbers come from an independent research firm, and the compliance gap SCANOSS describes is a structural feature of how agents operate, not a marketing claim.
94% of respondents said they lose productivity at least monthly because they have to analyze AI-generated code. Approximately one-third of teams use AI agents for comprehension and debugging only in straightforward codebases. ANALYSIS Together, these figures point toward a ceiling on agent utility in complex, mission-critical systems: the comprehension tax is not an edge case but a baseline condition.
What's next
Undo is proposing that AI be applied across more of the software-delivery cycle, giving debugging agents the ability to independently gather evidence and data to reach accurate conclusions, with runtime context as the foundation. SCANOSS's Earnie is structured as a continuous program with dated releases, positioning it for environments where code volume will keep rising. ◆ Both bets assume the same future: agent-generated code volume will continue to grow, and the enterprises adopting it will need new tooling layers that did not exist when humans wrote most of the code. The productivity story for AI coding is not wrong; it is incomplete, and the missing chapters are now being priced.