Original Post: https://www.linkedin.com/pulse/ai-sustainable-engineering-generate-fast-verify-thoroughly-yoon-hclqf/
[Notice / Disclaimer] All opinions expressed here are solely my own and do not represent the views of my employer. All references cited in this article are based on publicly available sources.
Preparing for What Lies Beyond Technical Debt: Comprehension Debt and Cognitive Debt
We are living through a period where the productivity and operational practices of software engineering are being rapidly redefined.
AI is transforming nearly every stage of development — code generation, testing, deployment, and operations. Smaller teams can now move faster and create greater impact than ever before, and the ownership of each individual engineer has become more important than ever.
I believe this shift represents an enormous opportunity for engineers.
At the same time, I think there are values we need to be more intentional about preserving in the age of AI. The ability to deeply understand and reliably operate what we build is becoming just as important as the ability to build it quickly.
AI Is Already a Powerful Engineering Accelerator
I recently had the opportunity to experience AI’s productivity gains firsthand while working on a cloud infrastructure resource optimization project.
Tasks that would have taken considerable time manually — resource inventory extraction, usage analysis, deprecation target identification, and migration priority assessment — were completed in just a few days with AI assistance, a process that would have otherwise taken several weeks.
This experience is becoming the norm across the industry. OpenAI reported that their engineering team wrote code in roughly 1/10th the time compared to manual development. Google disclosed that over 25% of their new code is now AI-generated. And 61.8% of developers are already incorporating AI into their development processes.
But there was an interesting observation.
Generation was fast, but verifying whether the results truly aligned with the system architecture and operational environment was still a human responsibility.
Reviewing the AI-generated analysis and recommendations one by one, I realized that what mattered most was not the ability to use AI tools, but a deep understanding of the system and hands-on operational experience.
Lightrun’s 2026 report shows a similar pattern. 43% of AI-generated code required additional debugging after deployment — even after passing QA and staging — and validating a single AI-generated fix required an average of three manual redeployment cycles.
AI provides an excellent starting point. But the final quality and accountability still rest with the engineer.
New Operational Challenges Are Emerging
We are all familiar with the concept of Technical Debt — deferred refactoring for faster delivery, insufficiently documented design decisions, and systems that grow increasingly complex over time.
As AI tools become ubiquitous, I’ve noticed that new types of challenges are becoming increasingly important alongside traditional technical debt.
Comprehension Debt
AI-generated code works correctly and passes all tests. But over time, the number of people on the team who can fully explain the internal workings of that code tends to decrease.
On the surface, everything looks fine. CI/CD passes, services run normally. But when the time comes to modify that system or analyze a failure months later, insufficient team understanding can dramatically increase the difficulty of maintenance.
Currently, 41% of commercial code is written with AI assistance (GitClear/Sonar 2026). While productivity has increased, code churn has doubled and code duplication has quadrupled. In the AI era, I believe the ability to explain why something works the way it does is becoming just as important as the ability to build it quickly.
Cognitive Debt
Another compelling concept is Cognitive Debt. Introduced by Appian CEO Matt Calkins in April 2026, this concept addresses a governance challenge that every organization accelerating AI adoption needs to consider.
AI has made it significantly easier to create new tools and automation. But over time, the number of systems an organization needs to manage and understand also grows rapidly. At some point, the speed at which a team can comprehend and govern its entire system landscape may matter more than the speed at which it can build new features.
According to CAST’s analysis of over 10 billion lines of code across 3,000 companies, the global technical debt burden amounts to 61 billion workdays of remediation effort. This isn’t a problem with AI — it’s a new operational challenge that arises precisely because AI is so powerful.
Engineering Competencies in the Age of Automation
In 1983, Lisanne Bainbridge introduced the concept of “Ironies of Automation.” The core idea is simple:
As automation increases, humans do less — but the moments when humans must intervene become the most difficult and exceptional situations of all.
And while automation handles the majority of cases, humans may gradually lose the opportunity to deeply understand how the system works internally.
I believe this concept is becoming even more relevant in the AI era. AI dramatically reduces repetitive work. But areas with high ambiguity and complexity — incident response, trade-off decisions, and analysis of unexpected operational issues — remain firmly in the human domain.
This is not a story about the limitations of AI. It’s about what competencies we need to intentionally develop so that we can maintain deep system understanding even as AI takes on more of the work.
Generate Fast, Verify Thoroughly
The direction that many engineering organizations are emphasizing converges on similar principles.
1. Leverage AI as an Accelerator
AI dramatically reduces repetitive, time-consuming tasks. Final design decisions, operational stability reviews, and cross-system trade-off analyses remain critical responsibilities of the engineer. The ability to verify and improve AI-generated outputs will become an increasingly important competency.
The AI-DLC (AI-Driven Development Life Cycle) framework published by AWS points in the same direction. Its “Collaborative by design” principle is particularly noteworthy — as AI does more, the human verification role should be intentionally built into the design.
2. Go Beyond “Does It Work?” to “Can I Explain It?”
This applies to more than just code. Architecture documents, design documents, analysis reports generated by AI — all the same. AI as a drafting accelerator is excellent. But the moment you put your name on the output, it becomes yours.
Whether AI created it or you wrote it yourself, being able to explain every line of a deliverable that carries your name would be truly impressive — and a path toward becoming a higher-caliber engineer.
A good benchmark is this:
“If I had written this from scratch myself, would I really have designed it this way?”
I believe the process of verifying until you can honestly answer “yes” to that question is what matters most.
3. Intentionally Accumulate Operational Experience
On-call rotations, incident response, Root Cause Analysis (RCA), and dependency mapping are not mere maintenance tasks — they are invaluable learning processes that deepen system understanding.
As Dr. Richard Cook observed, “Failure free operations require experience with failure.” In the AI era, the value of such experience may actually increase.
It may sound paradoxical, but it captures something essential about how operational expertise is built.
4. Invest in Tidying Up as Much as in Building
Creating new systems and automation is getting easier every day. But what may matter more in the long run is removing unnecessary resources, simplifying systems, and maintaining structures that the team can understand.
This is something I felt strongly during my own cloud resource cleanup work — AI was fast at generating the inventory, but assessing each resource’s dependencies and impact, and determining the deprecation sequence, required someone who understood the system.
5. Reinvest in Fundamentals
Data structures, algorithms, system design, networking — these fundamentals never go out of style. No matter how advanced AI tools become, the ability to structurally analyze problems during complex failures and make judgments from a whole-system perspective ultimately comes from fundamentals and experience.
As systems grow more complex, the value of these fundamentals only increases.
Sustainable Velocity
Using AI effectively is already becoming a baseline competency for every engineer. Beyond that, I believe engineers who combine AI proficiency with deep system understanding will create the greatest value going forward.
AI-powered rapid development is, of course, important and highly meaningful for any organization. I would add that the ability to reliably operate and maintain what has been built may determine long-term competitiveness. If a team can manage not only technical debt but also comprehension debt and cognitive debt, it will become an even stronger team over time — one that amplifies synergy while contributing to the broader organization.
This is not about cost reduction. It’s about Sustainable Velocity.
Closing Thoughts
We are in the middle of a truly significant inflection point.
AI is dramatically expanding what engineers can achieve, and this trend will only accelerate. I see this as an opportunity for engineers, not a threat.
Yet within this acceleration, certain things will become even more important: the commitment to deeply understanding our systems, the effort to accumulate operational experience, the culture of tidying up as much as we build, and the discipline of generating fast while verifying thoroughly.
I believe these will be tremendous assets in your continued growth as an engineer.
In an era of rapid change, the value of what doesn’t change only goes up.
"Speed may define the early winners in the AI era, but long-term success will belong to the teams that can sustain understanding alongside acceleration." — Sunsick Yoon
What's your thought?
No comments:
Post a Comment