Cognitive debt and intent debt are two new categories of software health risk that emerge when AI agents generate code faster than teams can understand it. Together with technical debt, they form Margaret-Anne Storey's Triple Debt Model — a framework for diagnosing where a system's health is degrading and why.
The Triple Debt Model
Proposed by Margaret-Anne Storey (University of Victoria) in a February 2026 paper and endorsed by Martin Fowler in his April 2 Fragments post, the model identifies three interacting types of debt:
| Debt Type | Lives In | Accumulates When… |
|---|---|---|
| Technical Debt | Code | Implementation decisions compromise future changeability |
| Cognitive Debt | People | Shared understanding of the system erodes faster than it's replenished |
| Intent Debt | Artifacts | Goals, rationale, and constraints are poorly captured or maintained |
Technical debt limits how a system can change. Cognitive debt limits how teams can reason about change. Intent debt limits whether the system reflects what was meant to be built.
Why This Matters for Agentic Teams
The classic software engineering assumption was that the engineers who built a system could understand it — the main risk was poor implementation (technical debt). AI coding agents break that assumption. An agent can produce thousands of lines of well-structured, working code that none of the team fully understands, at velocity that overwhelms any review cadence.
The causation chain runs in both directions:
- Intent debt causes cognitive debt: if rationale is never captured, new team members (human or AI) cannot form accurate mental models.
- Cognitive debt causes technical debt: developers who don't understand a system make poor implementation decisions, then generate more AI code to patch around them.
Storey's core concern: even if AI agents produce code that is easy to read, humans may have simply lost the plot — not understanding what the program is supposed to do, how their intentions were implemented, or how to safely change it.
Relationship to Other Practices
Cognitive and intent debt are the systemic, long-term risks that Context Engineering and Harness Engineering partially address at the tactical level:
- Context engineering ensures agents have the right information now but doesn't guarantee teams accumulate shared understanding over time.
- Harness engineering guards against individual agent mistakes but doesn't prevent the team from losing the thread on what the system is actually for.
- Intent capture (ADRs, BDD specs, structured AGENTS.md files) is the closest direct mitigation for intent debt.
Emerging Mitigation Practices
Several practices are gaining traction in 2026 as teams grapple with these new debt types:
For cognitive debt:
- Require that at least one human on the team fully understands each AI-generated change before it ships
- Allocate explicit time for walkthroughs where developers explain code they didn't write
- Treat "understanding" as a deliverable, not a byproduct of shipping features
For intent debt:
- Adopt intent-first workflows: capture design rationale in ADRs before AI generates code
- Write BDD-style specifications that encode purpose, not just behavior
- Invest in domain modeling sessions as a prerequisite for agentic tasks
- Use AI-assisted intent capture — tools that extract goals and constraints from meetings and conversations and store them as structured artifacts
For both:
- Build regular knowledge-transfer checkpoints into sprint cadences
- Treat onboarding and offboarding as forcing functions for rebuilding shared mental models
- Use AGENTS.md files as a live repository of system intent accessible to both humans and agents
Why Assess, Not Trial
The Triple Debt Model is currently a conceptual framework, not a measured practice. There are no established metrics for "cognitive debt" equivalent to cyclomatic complexity or DORA metrics. The paper appeared at TechDebt 2026 and ICSE 2026; practitioner blogs by Addy Osmani ("Comprehension Debt") and coverage by Simon Willison confirm the concept is resonating, but teams are still in the early stages of figuring out how to operationalize it.
The framework is worth understanding now because cognitive and intent debt compound quietly — by the time a team notices, they may have months of AI-generated code no one fully owns. The right response is to assess how much of this risk already exists in your systems and begin establishing intent-capture habits before velocity makes it unmanageable.
Key Characteristics
| Property | Value |
|---|---|
| Concept origin | Margaret-Anne Storey, University of Victoria, February 2026 |
| Endorsed by | Martin Fowler (April 2, 2026), Simon Willison (February 15, 2026) |
| Related concepts | Technical Debt, Comprehension Debt (Addy Osmani), Epistemic Debt |
| Mitigation practices | ADRs, BDD specs, intent-first workflows, AGENTS.md, mandatory understanding reviews |
| Academic venues | TechDebt 2026, ICSE 2026 |
Further Reading
- From Technical Debt to Cognitive and Intent Debt (arxiv 2603.22106) — Storey's original paper
- Martin Fowler Fragments, April 2, 2026
- Simon Willison coverage, February 15, 2026
- Comprehension Debt — the hidden cost of AI-generated code (Addy Osmani)
- Cognitive Debt: The hidden risk in AI-driven development (DX newsletter)
- Collaborative AI Engineering — the upstream practice that prevents alignment debt at the team level; addresses the organisational failure mode that cognitive and intent debt create when agents accelerate faster than teams can coordinate