From Technical Debt to Cognitive and Intent Debt: Rethinking software health in the age of AI

Keynote at Technical Debt Conference, Rio de Janeiro, April 13, 2026

Transcript

1. From Technical Debt to Cognitive and Intent Debt Rethinking software health in the age of AI Margaret-Anne Peggy Storey Technical Debt Conference - April 13, 2026
2. © M Storey How startups leverage vibe coding Around a quarter of those accepted into the accelerator use AI to write 95% or more of their code.
3. © M Storey “The SWE role will move ‘up the stack’ to be a bit more high level. The PM job will move ‘down the stack’ and a bit closer to code. We will meet in the middle and swap a lot of accountabilities” Ulloa, Butler, Haniyur, Miller, Amos, Sarkar, Storey: Product Manager Practices for Delegating Work to Generative AI: "Accountability must not be delegated to non-human actors" ICSE SEIP 2026. Thurs, 4:45 in Oceania V. Blurring of software team roles
4. © M Storey 4 Bringing software engineering and entrepreneur students together
5. © M Storey …gave us code that was clean and reusable and helped us progress a lot faster UI looked exactly how I had imagined it it taught me (a business student) so much… Fall 2025: Four weeks in….
6. © M Storey I spent a lot of time debugging small changes were hard, ended up scrapping everything it would have been quicker to just do it myself Fall 2025: Eight weeks in….
7. AI had accelerated the accumulation of technical debt … But it was cognitive and intent debt that paralyzed them © M Storey
8. Prototype → Product Journey Canvas . [ Prototype ] [ Product ] Mindset? Technical debt? Cognitive debt? Short term priorities Circle and rank the priorities you will focus on in the next sprint. Add missing priorities and connect related priorities. First paying users Validate ideas Evolvability Security Performance Integration Scalability Usability Observability Privacy Plan going forward Option A: Keep going, hope it holds out as is Risks? Mitigations? Option B: Keep but improve what you have Key improvements? Option C: Throw away, and start afresh What? How? Who? High Mark x for where you are now and y for where you want to be in next sprint. List the different types of technical and cognitive debt. Note if high / low for both. Choose (underline) 1 option. Add details as prompted for the selected option. © M Storey, Oct 2025
9. Prototype → Product Journey Canvas . [ Prototype ] [ Product ] Mindset? Technical debt? Cognitive debt? Short term priorities Circle and rank the priorities you will focus on in the next sprint. Add missing priorities and connect related priorities. First paying users Validate ideas Evolvability Security Performance Integration Scalability Usability Observability Privacy Plan going forward Option A: Keep going, hope it holds out as is Risks? Mitigations? Option B: Keep but improve what you have Usability, payments, scalable for >1K users. Option C: Throw away, and start afresh What? How? Who? High Mark x for where you are now and y for where you want to be in next sprint. List the different types of technical and cognitive debt. Note if high / low for both. Choose (underline) 1 option. Add details as prompted for the selected option. © M Storey, Oct 2025 2 1 3 4 Architecture (not easy to add features), brittle (buggy when we make any changes) One person understands it well and how make changes but others do not X Reflection: Not as far along as we hoped, but we pivoted. But we know what we want now and one of us understands the code well so we can move to implementing the features that will be more robust Y
10. © M Storey “I've been experimenting with prompting entire new features into existence without reviewing their implementations… I've found myself getting lost in my own projects. I no longer have a firm mental model of what they can do and how they work” – Simon Willison, https://simonwillison.net/2026/Feb/15/cognitive-debt/ Not just in my Startup course…
11. © M Storey This talk is a summary of a conversation I’ve had with practitioners about cognitive and intent debt, and proposed triple debt model1 Acknowledgements: Arty Starr, Adam Tornhill, Keith Mann, Kent Beck, Dave Thomas, and many others… 1. From Technical Debt to Cognitive and Intent Debt: Rethinking Software Health in the Age of AI. M. Storey. To appear ACM Queue 2026 https://margaretstorey.com/blog/2026/02/18/cognitive -debt-revisited/
12. © M Storey Technical debt : a familiar concept
13. Sometimes intentional, trading quality for speed Technical debt limits optionality (Kent Beck) Measurable, observable, refactorable in principle Lives in the code and design TECHNICAL DEBT Not always bad – Fowler’s quadrant
14. TECHNICAL DEBT IN THE AI ERA The jury is still out AI MAKES IT WORSE Poor quality generated code. Semi-approval already happening. AI CAN HELP Automated refactoring, test generation, code explanation, skills. CODE MAY DISAPPEAR Some argue code may be hidden infrastructure managed by agents. Code is “accidental complexity”. THE DEEPER RISK AI is a multiplier, and right now it's multiplying the debts we're worst at managing.
15. © M Storey Cognitive debt : the invisible layer Technical debt lives in the code. Cognitive debt lives in people.
16. Peter Naur (1985): a program is a theory held in developers’ minds, not just source code. What is software? 2. DESIGN JUSTIFICATION Why structural decisions were made: data models, integration points, architecture choices. 1. WORLD-TO-PROGRAM MAPPING How real-world concepts map to program structure. What's included, what's left out. 3. MODIFICATION CAPACITY Can the team reason about safe change paths? (Biggest risk with AI-generated code.)
17. THEORETICAL GROUNDING: DISTRIBUTED COGNITION Software isn't understood by individuals: it's understood by teams PEOPLE Developers, PMs, agents ARTIFACTS Code, docs, tests, specs INTERACTIONS Reviews, retros, conversations Understanding is distributed across all three. Today, teams include both people and AI agents.
18. 18 Cognitive debt A project-level property reflecting the erosion of shared understanding of the software across a software team over time.1 Persistent gaps in shared mental models, transactive memory, common ground Different from momentary cognitive load and individual comprehension debt2 at the level of the source code With AI, cognitive debt accumulates faster due to cognitive surrender3 2. Alakmeh, Anderson, Jackson, Vaz Pereira, Akirmak, Estey, Prikladnicki, van der Hoek, Storey & Fritz, Grasping AI Reliance in Program Comprehension and Coding through the AIRELI Persona Taxonomy. ICPC 2026, 3. Shaw, S. D. & Nave, G. “Thinking-Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender.” Available at SSRN 6097646 (2026).
19. The software enchiridion and cognitive debt “We tell ourselves lots of half-truths and, in this age of AI assisted coding… Not the brazen lie of complete competance. Not the comic lie of the demo that “mostly works. The polished, rational lie borne of urgency: We understand it well enough.” “In software, we call this “moving fast.” In reality, we are borrowing against cognition…. Eventually pay interest in confusion” – Russell Miles (response to Cognitive Debt post4) 19 4. https://www.softwareenchiridion.com/p/on-cognitive-debt-and-the-care-of
20. Diagnosing cognitive debt INDIVIDUAL SIGNALS Confusion about unexpected behavior.5 Cognitive fatigue trying to figure it out.5 Reluctance to modify code written by AI. TEAM SIGNALS People ask each other basic questions. Slow onboarding. Low bus factor.6 No awareness of who knows what. 5. A. Starr & M.A. Storey. Theory of Troubleshooting: The Developer's Cognitive Experience of Overcoming Confusion. To appear ACM TOSEM. 2026 6. Tornhill, A. 2015. Your Code as a Crime Scene. Pragmatic Programmers.
21. Preventing & Mitigating Cognitive Debt Postmortems and Retrospectives for reflecting and regrouping when things break System Walkthroughs for the purpose of shared understanding of architecture Habitat Thinking4 design the environment for (re)building shared understanding across the team Design for Troubleshooting5 ensure “diagnostic legibility”4 is preserved as developers lack familiarity with code they didn’t write but may need to troubleshoot
22. 22 Should we measure cognitive debt? Two proposals: COGNITIVE DEBT SCORE8 Five questions to ask in a retro: 1. Explain modules without looking at code? 2. Can you debug at 2am without AI? 3. Do you know why this implementation? 4. Long time to onboard someoneone new? 5. Is code maintainable if AI goes away? Yes < 3/5 → Cognitive Debt! FEATURE COMPREHENSION SCORE7 AI-generated quiz on Naur's 3 layers: 1. world-program mapping 2. design justification 3. modification capacity scored at team level per feature, in retros. 7. Leonid Sokolovskiy: https://levelup.gitconnected.com/we-can-measure-how-fast-we-ship-can-we-measure-how-well-we-understand-wha t-we-built-a5ad2749c8fa 8. Adham Sersour: https://www.linkedin.com/feed/update/urn:li:activity:7442198356598239232/ But caution is needed when we create new metrics as we may damage what we can’t see but value…
23. © M Storey Intent Debt : the hidden layer Technical debt lives in the code. Cognitive debt lives in people. Intent debt lives in the absence of artifacts.
24. Rahul Garg’s AI frustration loop “AI generates solutions misaligned with existing architecture” → because the architecture's why was never externalized “Developers spend significant time on post-generation editing” → paying the interest on that missing rationale “Context established early in a conversation is lost as session lengthens” → intent lives nowhere persistent “Quality varies depending on which team member is prompting” → because intent is locked in people's heads, not artifacts 9. R. Garg, Patterns for Reducing Friction in AI-Assisted Development, https://martinfowler.com/articles/reduce-friction-ai/ April 8, 2026
25. Intent debt The absence or erosion of externalized rationale, goals, and constraints that guide how humans and AI agents evolve the system. It lives in artifacts, or more often, in the absence of them.1 “AI is a loudspeaker, not an asymmetric force. A team with strong intent practices finds AI amplifies clarity. A team with weak practices finds AI fills gaps with plausible guesses.” — Sandro Ponticelli (comment to Triple Debt Model post)
26. Diagnosing intent debt Changes don't line up and no one can explain why, users are unhappy Agents struggle, they make plausible but wrong decisions, use more tokens Nothing is written down, forgotten requirements, constraints
27. Mitigating Intent Debt DECISION AND RATIONALE RECORDS Architectural Decision Records capture what was decided and why. Domain-Driven Design makes domain intent explicit before it is encoded EXECUTABLE INTENT BDD specifications and tests to verify purpose may show if the system may drifts from its intent CONTEXT ARTIFACTS Context and harness engineering (skills, agent instructions, and playbooks) produce feedforward and feedback guidance for both humans and agents
28. A Return to Fundamentals: McLuhan’s Retrieval “wow, I've seen this several times in my career, didn't know it had a name. ” – Alistair Cockburn, LinkedIn post Storey, Russo, Novielli, Kobayashi, and Wang. A Disruptive Research Playbook for Studying Disruptive Innovations. TOSEM 2024.
29. © M Storey The Triple Debt Model Three interacting layers of software health
30. TRIPLE DEBT MODEL Three layers, three locations1 TECHNICAL DEBT Lives in the code. Observable, measurable, refactorable. Focus of this conference and still critical. COGNITIVE DEBT Lives in people. Erosion of shared understanding across a team over time. Invisible until it paralyzes. INTENT DEBT Lives in artifacts, or their absence. Missing rationale that humans and agents need to evolve the system safely. AI tools are very good at producing code and measuring technical debt. Very few are optimized for improving understanding or preserving intent. This may be where the real risk lies.
31. How the three debts interact COGNITIVE → INTENT When shared understanding erodes, capturing intent becomes harder because no one knows what rationale to document. TECHNICAL → COGNITIVE Messy code is harder to build shared understanding around. Poor architecture fragments the team's mental model. INTENT → TECHNICAL Can the team reason about safe change paths? (Biggest risk with AI-generated code.)
33. “While I’m getting a bit bemused by debt metaphor proliferation, this way of thinking does make a fair bit of sense. The article includes useful sections to diagnose and mitigate each kind of debt. The three interact with each other, and the article outlines some general activities teams should do to keep it all under control” – Martin Fowler, Fragments Newsletter April 2, 2026
34. Extending Fowler’s Technical Debt Quadrant with Cognitive & Intent Debt “We have no time to understand or document intent” “We will understand it later, and document our decisions then” “Can anyone predict what happens if we make this change?” “We must rollback and recreate what we decided and implemented”
36. © M Storey What about Startup Studio Fall of 2026? Make intent explicit earlier Make sure they use engineering fundamentals Treat understanding as a deliverable
37. Created with Claude (one prompt!)
39. FUTURE WORK Possible next steps Ensure system is a good habitat Tools may not be enough to decrease frustration, fatigue and burnout. Cognitive debt isn't just a project risk; it's a physiological one (Starr & Storey) How to balance the three types of debt Impact on project outcomes? Acceptable amounts of debt for sustainability? What tradeoffs are ok? How to reason about these three types of debt? Can we measure and not introduce harmful metrics? Can we sense them without measurement? Other types of debt? Social debt, process debt, org. debt, skill debt, community debt… What role will AI play in causing and mitigating all of these?
40. Why this matters Technical debt makes systems harder to change Cognitive debt makes systems harder to understand Intent debt makes it hard for humans and agents to know what the system is for
41. 41 mstorey@uvic.ca http://margaretannestorey.com https://www.linkedin.com/in/margaret-anne-storey/ Acknowledgements: Arty Starr, Adam Tornhill, Keith Mann, Kent Beck, Dave Thomas, Martin Fowler, and others…