Technical, cognitive, and intent debt

· coding ai ai-agents · Source ↗

TLDR

  • Margaret-Anne Storey’s framework splits system health into three debt layers: technical (code), cognitive (people), and intent (artifacts), with AI making intent debt the most dangerous.

Key Takeaways

  • Intent debt lives in artifacts: when goals and constraints are poorly captured, neither humans nor AI agents can reliably evolve the system.
  • Shaw and Nave propose LLMs as System 3 cognition; “cognitive surrender” is passive reliance on AI output, distinct from deliberate “cognitive offloading.”
  • Ajey Gore: when agents make coding free, verification becomes scarce. “Correct” is thousands of shifting, context-dependent definitions across 900 microservices.
  • Org restructuring follows: a 10-engineer feature team becomes 3 builders and 7 people designing acceptance criteria, test harnesses, and monitoring outcomes.
  • Fowler sees humans growing a domain-specific Ubiquitous Language with LLMs as the remaining creative act, naming structure that maps cleanly to the problem.

Hacker News Comment Review

  • Commenters split on the abstraction-layer argument: if Assembly-to-Python already created cognitive and intent debt by this definition, the framework may describe a universal tradeoff rather than an AI-specific crisis.
  • Legacy migration generated real tension: some found LLM-assisted boundary design plus test-parity harnesses practical for modernization; others argued you cannot “own” code you never organically developed, especially for databases or stateful systems.
  • The cognitive surrender paper from Wharton drew pointed skepticism: commenters noted the paper is largely AI-generated and not yet peer-reviewed, which undercuts its authority as a source on uncritical AI reliance.

Notable Comments

  • @hibikir: LLMs can be prompted toward minimal changes and deduplication passes; the instincts of a senior dev are available but not defaults.
  • @ryanisnan: Concrete counterexample: Claude reused existing DB schema models for a new logical concept rather than adding new ones, forcing consumer-side gymnastics.
  • @meander_water: Flags that the Wharton cognitive surrender paper is “entirely AI generated” and unreviewed, calling the irony hard to ignore.

Original | Discuss on HN