AI is making developers faster. That part is real. But speed has a hidden cost.
The same longitudinal study that found 32% faster PR review cycles also found that worsened developer experiences nearly doubled — from 14% to 27% — over the same period. The METR RCT found developers felt 20% faster while actually being 19% slower.
Developers are shipping faster. And quietly, something is eroding underneath.
This repo is the companion resource for the lightning talk "Cognitive Debt in the Age of Agentic AI". It covers the research behind the talk, all six failure modes in depth, and concrete fixes you can apply this week.
These aren't independent problems. They form a reinforcing feedback loop:
graph LR
A(["🔍 Opacity"]) --> B(["👤 Accountability Gap"])
B --> C(["⚖️ Trust Miscalibration"])
C --> D(["🔀 Context Switching"])
D --> E(["🌊 Flow Disruption"])
E --> F(["📉 Skill Atrophy"])
F -->|"feeds back"| A
style A fill:#e74c3c,color:#fff
style B fill:#e67e22,color:#fff
style C fill:#f1c40f,color:#000
style D fill:#2ecc71,color:#fff
style E fill:#3498db,color:#fff
style F fill:#9b59b6,color:#fff
Each mode feeds the next. Fixing comprehension (Opacity) is upstream of all the others.
| # | Mode | One-liner | Read |
|---|---|---|---|
| 01 | Opacity | You ship code you don't understand | failure-mode · mitigation |
| 02 | Accountability Gap | Responsibility diffuses when AI-generated code fails | failure-mode · mitigation |
| 03 | Trust Miscalibration | Your confidence in AI output exceeds its actual reliability | failure-mode · mitigation |
| 04 | Context Switching | Managing agents overwhelms working memory | failure-mode · mitigation |
| 05 | Flow State Disruption | Supervising an agent is not the same as writing code | failure-mode · mitigation |
| 06 | Skill Atrophy | Delegation erodes the skills you stop practicing | failure-mode · mitigation |
Modes 01, 03, and 06 were covered in the talk. All six are in the repo.
If you have 5 minutes: Read overview.md — the full picture in one page.
If you want the slides: Available on request — open an issue or reach out directly.
If you want the fixes now: agents-md-snippets.md — all six AGENTS.md blocks, ready to copy.
If you want the evidence: sources.md — annotated bibliography with links to every cited study.
| Failure Mode | Guardrail |
|---|---|
| Opacity | Comprehension gate — can't explain it to a colleague? Don't merge it. |
| Trust Miscalibration | Measure, don't feel — track actual cycle time. Don't trust the feeling of speed. |
| Skill Atrophy | Solve-first protocol — attempt the problem yourself before opening the chat. |
- GitHub RCT (arXiv:2302.06590) — 55% faster task completion, 95 professional developers
- Longitudinal study (arXiv:2509.19708) — 300 engineers, 1 year; worsened DX nearly doubled (14% → 27%)
- METR RCT 2025 (metr.org) — felt 20% faster, were 19% slower; 39-point perception gap
- Anthropic RCT 2025 (anthropic.com) — 17% lower comprehension scores; biggest gap: debugging
- Stack Overflow 2025 (survey.stackoverflow.co/2025) — 84% adoption, only 3% highly trust AI output accuracy
Full annotated bibliography → sources.md
01-opacity/ — You ship code you don't understand
02-accountability-gap/ — Responsibility diffuses when AI code fails
03-trust-miscalibration/ — Your confidence exceeds AI's actual reliability
04-context-switching/ — Managing agents overwhelms working memory
05-flow-state-disruption/ — Supervisory work is not deep work
06-skill-atrophy/ — Delegation erodes the skills you stop practicing
agents-md-snippets.md — All six AGENTS.md blocks in one place
overview.md — Full picture in one page
sources.md — Annotated bibliography
Each mode directory contains failure-mode.md (what it is, why it happens, evidence, costs) and mitigation.md (baseline fixes, AGENTS.md snippet, experimental tools).
Research synthesis and presentation materials © 2025 The LEGO Group. Source studies retain their original licenses — see sources.md for links.