Intent Debt in Software Health

• Intent debt is defined as the erosion or absence of explicit rationale, goals, and constraints captured in external artifacts that guide system evolution.
• It is a critical component of the triple debt model, linking technical, cognitive, and intent debt to misalignment between system behavior and stakeholder expectations.
• In AI-assisted development, intent debt emerges as gaps between latent source intent and encoded representations, leading to prompt brittleness and governance challenges.

Searching arXiv for recent and relevant papers on “intent debt” and adjacent concepts to ground the article in current literature. arxiv_search(query="all:\"intent debt\" OR all:\"intent-centric software engineering\" OR all:\"intent signal theory\" OR all:\"technical debt\" software health AI", max_results=10, sort_by="submittedDate") Intent debt denotes the absence or erosion of explicit rationale, goals, and constraints that guide how a system evolves. In software-health research, it is the debt that lives in externalized knowledge artifacts rather than in code or in people’s heads; in adjacent work on human-AI interaction, it is formalized as unrecovered intent, the weighted portion of source intent that is not encoded in the carrier presented to the model. Across these usages, the common concern is not merely implementation quality but the persistence, recoverability, and governability of the system’s “why” (Storey, 23 Mar 2026, Peng, 24 May 2026).

1. Emergence and scope of the concept

The notion of intent debt emerged against a broader expansion of the debt metaphor beyond source code. Earlier work had already argued that technical debt can be observed directly in maintenance behavior through “debt-prone bugs,” specifically tag bugs, reopened bugs, and duplicate bugs (Xuan et al., 2017). Parallel work on agile development identified non-technical debt in process, social, people, and organizational forms, but did not define a subtype named intent debt (Ahmad et al., 1 Sep 2025). Within this widened debt landscape, intent debt was introduced explicitly as part of a triple debt model for software health (Storey, 23 Mar 2026).

In that model, intent debt is the “forgotten layer.” The term is used for deficits in explicit, durable, consultable representations of system purpose and rationale. This distinguishes it from debt that resides in implementation structure and from debt that resides in human understanding alone. A plausible implication is that intent debt becomes visible only when a project asks not whether the code runs, but whether the organization can still explain what the system is for, what constraints govern it, and why it evolved as it did.

Current literature also uses the term in a second, more computational sense. In human-AI interaction, intent debt refers to the gap between latent source intent and the prompt carrier, especially when omitted private dimensions cannot be faithfully recovered later (Peng, 24 May 2026). The two usages are closely related: both concern losses in explicit intent representation, though one is oriented toward software evolution and the other toward intent encoding and recovery.

2. Intent debt in the triple debt model of software health

The triple debt model distinguishes three interacting debt types: technical debt in code, cognitive debt in people, and intent debt in artifacts. Technical debt lives in code and architecture; cognitive debt lives in shared human understanding; intent debt lives in externalized knowledge artifacts such as requirements documents, architectural decision records, implementation plans, acceptance tests, specifications, domain models, and other context artifacts for AI-assisted development (Storey, 23 Mar 2026).

Intent debt is defined as “the absence or erosion of explicit rationale, goals, and constraints that guide how a system evolves.” It is not treated as vague missing documentation. The emphasis is on whether objectives, constraints, and decision rationale are clearly articulated and captured in artifacts that both humans and AI systems can consult. The model therefore assigns intent debt to the external memory of a software organization rather than to code quality or individual expertise.

The model frames software health as the alignment of three layers: intent, code, and understanding. When these layers drift apart, different debts accumulate and amplify one another. Intent debt can cause cognitive debt because developers cannot form accurate mental models when the system’s purpose is not articulated. Cognitive debt can in turn cause intent debt because developers who do not understand the system cannot properly externalize specifications and decisions. Cognitive debt can also cause technical debt, while technical debt can amplify cognitive debt by making the code harder to reason about. The literature treats the reinforcing relationship between cognitive debt and intent debt as especially important.

The primary accumulation mechanisms are decisions made informally and never recorded, requirements that drift over time, incomplete or outdated specifications, goals and constraints known only to a few people, AI-generated code that bypasses the human process of articulating intent, and overreliance on generated outputs without preserving decision rationale. The recommended temporal discipline is explicit: intent should ideally be captured at the moment key decisions are made, because reconstructing it later can be difficult or impossible.

The practical failure mode is alignment loss. If intent is not captured, behavior can drift from stakeholder expectations, teams may build the wrong thing, AI agents may optimize for the wrong objectives, and future changes become riskier because the rationale is unavailable. In this formulation, intent debt is not an auxiliary documentation concern; it is a first-order determinant of whether the system still reflects intended goals.

3. AI-assisted development and the acceleration of intent debt

Generative AI changes the economics of software creation in a way that sharpens intent debt. One line of work argues that LLM-assisted development amplifies traditional technical debt categories such as code debt, design debt, and documentation debt, while also introducing prompt debt, fast-integration debt, governance debt, data debt, ethical debt, and provenance debt (Ehsani et al., 11 Jun 2026). Prompt debt is defined as debt from unclear or undocumented prompts, and fast-integration debt as risks from rapidly integrating LLM-generated outputs without proper validation, leading to unstable foundations. The causal chain is explicit: faster generation encourages immediate integration, superficial validation leaves hidden bugs and architectural mismatches in place, and long-term maintenance cost increases.

A related line of work describes a transition “from code-centric production toward intent-centric human-agent work,” in which natural language, repository context, tools, tests, and governance increasingly shape delivery (Cruz, 10 May 2026). Under this framing, the scarce work shifts upstream and sideways: intent specification, context curation, supervision, verification, governance, and evidence production become more consequential than isolated code authorship. The unit of engineering shifts from a human editing a file to a human-agent system operating across requirements, code, tests, build pipelines, scanners, deployment workflows, and telemetry.

These arguments converge on the same structural risk. LLMs make plausible code cheap, but they do not remove architecture knowledge, security reasoning, change management, or long-term ownership burdens. When intent is vague, context is thin, and validation is weak, generated code can satisfy a surface-level request while missing deeper design intent, governance constraints, or future changeability needs. This suggests that intent debt in AI-assisted development is not confined to missing rationale artifacts; it also includes underinvestment in the specification, context, and evidence needed to keep machine-generated changes aligned with system purpose.

The literature therefore links intent debt to several recognizable downstream symptoms: code that “runs” but does not align with intended architecture, prompt brittleness, non-determinism, poor reproducibility, unclear provenance, and what one source describes as “the growing gap between the code that exists and the code we understand” (Ehsani et al., 11 Jun 2026). In this sense, intent debt becomes one mechanism by which speed-focused adoption creates hidden maintenance and accountability liabilities.

4. Computational formalization in human-AI interaction

Intent Signal Theory formalizes the intent layer that prompt-centric models omit. It distinguishes four objects routinely conflated in LLM interaction: latent source intent $I^*$, observable intent proxy $I$ or $\hat{I}$, encoded carrier $P$, and model output $O$. The observable proxy is represented as a finite weighted set of dimensions,

$$I = \{(e_i, W_i)\}, \qquad W_i \ge 0, \qquad \sum_i W_i = 1.$$

The prompt carrier encodes dimensions through a mask $m_i \in \{0,1\}$, and encoding loss is defined as

$L_{\text{enc}} = 1 - \sum_i W_i m_i.$

This is the formal version of intent debt in the theory: the weighted portion of intended meaning absent from the carrier (Peng, 24 May 2026).

The theory then separates structural recovery from fidelity recovery. Structural recovery asks whether the output fills the expected slot for a dimension; fidelity recovery asks whether it reproduces the user’s actual intended content. The weighted aggregates are

$s\text{-}ICM_w = \sum_i W_i r_i$

and

$f\text{-}ICM_w = \sum_i W_i f_i,$

with intent drift defined as

$I$0

The central observation is the structural-fidelity split: outputs may look structurally complete while remaining semantically wrong relative to the user’s actual goal.

A further distinction is between public and private intent dimensions. Public dimensions are predictable from the model’s prior; private dimensions are specific to the user or context and cannot be reliably recovered at the specificity needed for fidelity unless they are explicitly encoded. This distinction underwrites the Theorem of Irreversible Intent Loss. Under single-turn interaction, omitted encoding of a private dimension, and non-recoverability of the intended value from the prior, no decoder operating only on the reduced carrier and model prior can recover that value beyond generic substitution. The information-theoretic statement is

$I$1

and under the privacy assumption,

$I$2

therefore

$I$3

The empirical program associated with the theory reports several regularities. Structured intent encoding using 5W3H/PPS improved alignment across 540 outputs; the structured-encoding advantage held across Chinese, English, and Japanese, across six model families, and across 2,160 outputs. In the measurement layer, outputs often exhibited high structural recovery but low fidelity recovery: 25.7% of outputs in Chinese and 58.6% in English fell in the split zone. Across the ablation set, 31.5% of cells were public-regime and 68.5% were private-regime. Human raters gave split-zone outputs a mean GA of 3.12, while LLM judges gave them 5.0; human-LLM agreement on dimensional $I$4 was 0.695, compared with 0.251 on holistic GA. The theory’s practical conclusion is that prompt engineering is better understood as intent-protocol design than as surface text optimization.

5. Governance, mediation, and the relocation of hidden debt

A distinct but adjacent response to intent debt appears in intent-driven computing. Here an intent is defined as a finite, structured data value

$I$5

representing a proposed effectful operation. Programs produce intents rather than directly executing effects; a governed runtime examines each intent against a decidable policy language, records every decision in a tamper-evident ledger, and only then realizes the effect. The runtime sequence is fixed: program produces an intent, governance interpreter checks policy, the decision is appended to a tamper-evident ledger, allowed effects are realized, denied effects are blocked, and escalated effects wait. The key structural claim is that “No alternative path to effects exists in the language semantics” (McCann, 21 May 2026).

This model yields several emergent properties relevant to intent debt. Audit Completeness states that every effectful action taken by an intent-driven program has a corresponding decision record in the ledger. The ledger is hash-chained,

$I$6

so audit can be independently verified. Because every intent produces a decision record, the ledger is, by construction, an event store. Historical intent streams can also be replayed for governance simulation, allowing counterfactual analysis of alternative policy deployments.

The paper does not define intent debt as a formal construct, but it explicitly motivates the model as a way to make effects explicit, inspectable, replayable, and auditable. This suggests a governance-oriented interpretation of intent debt: hidden operational and audit debt is reduced by forcing all effectful interaction through explicit intent values. The cost is not eliminated; it is relocated into policy design, ledger maintenance, trusted runtime components, context reconstruction, and mediation overhead. The reported overhead is measurable: policy evaluation is sub-microsecond to about 1–2 $I$7, SHA-256 hash computation is sub-microsecond, ledger append is hundreds of microseconds, and total governance is about 535 $I$8 p50. The trusted computing base includes the BEAM VM, the OCaml extraction pipeline, the native interface bridge, the OS, and hardware. The model therefore turns hidden governance debt into visible, structurally enforced intent-management burden.

6. Management, indicators, and conceptual boundaries

The software-health literature treats intent debt as manageable, but only if teams preserve intent as a first-class artifact. Recommended practices include treating understanding as a deliverable, using intent-first workflows, capturing intent early when key decisions are made, maintaining living artifacts, and resisting the use of AI merely to generate documentation that looks explanatory without building real human understanding. Suggested monitoring spans the three layers together and may include onboarding time, knowledge concentration metrics, requirements coverage analysis, and audits comparing documented intent with actual behavior (Storey, 23 Mar 2026).

Several adjacent literatures are relevant but conceptually distinct. The survey on self-fixed technical debt studies whether practitioners consciously and often repay debt they themselves introduced, and finds that a majority addresses their own debt consciously and often; this is evidence about intentional repayment behavior, not a formal definition of intent debt (Tan et al., 2021). The debt-aversion survey module measures a financial preference parameter $I$9, where $\hat{I}$0 indicates debt aversion; the paper explicitly states that it does not directly study a concept formally named intent debt (Albrecht et al., 2022). The non-technical debt white paper is directly concerned with process, social, people, and organizational debt and is only indirectly relevant through its discussion of unclear roles, misaligned workflows, lack of shared vision, and poor synchronization (Ahmad et al., 1 Sep 2025).

A persistent misconception is to reduce intent debt to ordinary documentation debt. The literature is more specific. Intent debt concerns the durable capture of rationale, goals, constraints, and decision records in forms that remain usable for both humans and machines. Another misconception is to treat intent debt as eliminated once code is generated or an effect is mediated. The surveyed work instead suggests a stricter conclusion: intent debt is reduced only when intent is sufficiently externalized, recoverable, auditable, and governable across the full lifecycle of change.