Intent Engineering: Key Foundations

• Intent Engineering is a discipline that defines and captures high-level goals, constraints, and preferences to drive system task execution.
• It integrates methodologies from HCI, networking, and AI to separate user aims from system operations, thereby enhancing control and evaluation.
• Applications include structured prompt construction, intent signal theory, and lifecycle modeling to improve task quality, safety, and governance.

Searching arXiv for the cited Intent Engineering and related papers to ground the article. Intent Engineering (IE) denotes the end-to-end process or discipline through which a system elicits, represents, operationalizes, assures, and evaluates intentions that drive task execution. In the HCI formulation, IE separates “what the user aims to achieve” from “how the system performs the steps,” emphasizing high-level goals, constraints, preferences, and iterative co-specification with AI; in retrieval it makes intent a first-class control signal for a dense retriever; in networking it covers modeling, specifying, validating, decomposing, implementing, operating, and governing intents across their lifecycle; and in enterprise agent systems it encodes organizational goals, priorities, values, trade-off hierarchies, ethical principles, and decision boundaries into the infrastructure on which agents operate (Ding, 2024, Pan et al., 2023, Dzeparoska et al., 2024, Vishnyakova, 10 Mar 2026).

1. Definitions and conceptual boundaries

IE is defined in multiple, partially overlapping ways. The HCI and interface-centered account defines it as “the end-to-end process through which a system elicits, represents, operationalizes, and evaluates user intentions to drive task execution,” with explicit distinctions among intent capture, intent representation, intent operationalization, and intent evaluation (Ding, 2024). In Intent-Based Networking, IE is the lifecycle discipline of “modeling, specifying, validating, decomposing, implementing, operating, and governing intents,” with runtime assurance and conflict resolution treated as integral rather than auxiliary concerns (Dzeparoska et al., 2024). In corporate multi-agent architecture, IE is “the practice of encoding corporate goals, priorities, values, trade-off hierarchies, ethical principles, and decision boundaries into the infrastructure on which agents operate” (Vishnyakova, 10 Mar 2026). In intent-centric software engineering, humans increasingly specify “desired outcomes, constraints, quality attributes, policies, and acceptance evidence,” while agents generate, test, repair, and explain candidate implementations (Cruz, 10 May 2026).

These accounts converge on a common boundary: intent is not merely a prompt string, a command sequence, or a credential. The paper on Intent Signal Theory separates latent source intent $I^*$, observable intent proxy $I$, encoded carrier $P$, and model output $O$, arguing that prompt surface alone is an incomplete proxy for the user’s source intent (Peng, 24 May 2026). The IntentTagger work similarly distinguishes monolithic prompting from “intent tags,” that is, small, atomic conceptual units used for granular and non-linear micro-prompting in co-creation workflows (Gmeiner et al., 26 Feb 2025). This suggests that IE is broader than prompt engineering: it includes prompt construction, but also includes decomposition, provenance, policy, workflow design, and fidelity assessment.

A related conceptual boundary concerns context. “Context Engineering” is described as the design and management of the informational environment in which an agent acts, whereas IE supplies the purposive layer that determines “what outcome to pursue and what to sacrifice” (Vishnyakova, 10 Mar 2026). The practical implication is that adequate context does not, by itself, guarantee alignment with strategic or user intent.

2. Formal representations and lifecycle models

A canonical formalization treats an intent as $i \in I$ with attributes $i=(G,C,P)$, where $G$ are goals, $C$ constraints, and $P$ preferences; with context $c$ and model or pipeline $I$0, output generation is written as $I$1 (Ding, 2024). The same formulation defines a multi-criteria utility

$I$2

with refinement driven by feedback through $I$3, or equivalently by a Bayesian update over latent intent hypotheses,

$I$4

Within this lifecycle, capture elicits goals and constraints, representation structures them with context and provenance, operationalization maps them to executable plans, and evaluation assesses relevance, faithfulness, diversity, novelty, and task completion (Ding, 2024).

Intent Signal Theory introduces a different formal layer centered on transmission and recovery. It defines the chain

$I$5

where $I$6 is latent source intent, $I$7 is the observable proxy, $I$8 the carrier, and $I$9 the output (Peng, 24 May 2026). The proxy is decomposed into weighted dimensions $P$0 with $P$1, and encoded via a binary mask $P$2. Observable encoding loss is

$P$3

Recovery is then separated into structural slot-filling and fidelity to the intended content via

$P$4

with intent drift defined as $P$5. A central result, the Theorem of Irreversible Intent Loss, states that private intent absent from the carrier cannot be recovered beyond generic substitution under the single-turn, carrier-only assumptions specified in the theory (Peng, 24 May 2026).

Other subfields operationalize intent differently. I3, for retrieval, conditions dense retrieval on an instruction-derived introspected intent embedding injected into the query pathway while freezing the base retriever, thereby separating semantic matching from task-specific relevance semantics (Pan et al., 2023). IDIL, for imitation learning, models expert behavior with a time-varying latent intent $P$6, an intent-conditioned policy $P$7, and intent dynamics $P$8, then alternates intent inference and intent-aware policy learning (Seo et al., 2024). These formulations are not identical, but they share the assumption that intent is a control variable whose representation affects both execution and evaluation.

3. Interaction patterns and interface design

In interface research, IE is studied across a task spectrum rather than as a single interaction style. One design space spans intent granularity, explicitness versus implicitness, constraint richness, guidance level, iteration loops, provenance, controllability, ambiguity handling, and feedback mechanisms (Ding, 2024). The same work probes three representative task types: fixed-scope content curation through news headline generation, atomic creative tasks through cross-domain analogy generation, and complex interdependent tasks through exploratory visual data analysis. The interaction loops differ accordingly: single-shot textbox input often suffices for fixed-scope curation; select–rate–revise loops become important in creative tasks; and complex analysis requires continuous exploration, provenance management, and sensemaking on a large canvas (Ding, 2024).

IntentTagger pushes this interactional account toward micro-prompting. Its “Intent Tags” are atomic units such as concept tags and reference tags, organized into Narrative, Visual Style, and Content Sources groups, with deck-level and slide-level scope, tag grounding from text or from slides, drop-down alternatives, previews, and an “opposite” slider for exploring spectra such as Modern versus Traditional typography (Gmeiner et al., 26 Feb 2025). The reported design implication is that users can express intent across varying levels of ambiguity without repeatedly reconstructing a monolithic prompt. This supports “meta-intent” reflection: users discover missing constraints or preferences while manipulating persistent intent objects rather than rewriting instructions from scratch.

A complementary line treats intent perception itself as an interface problem. UI-JEPA predicts intent from onscreen interaction sequences using masked feature prediction over UI videos and a lightweight decoder, targeting low latency, privacy preservation, and on-device deployment (Fu et al., 2024). IEA, a conversational image editing agent, operationalizes intent through 16 explicit editing tools with integer parameters in $P$9, transparent edit traces, iterative refinement, and an Image-Summary task that converts past edits into reusable preference descriptions (Zhu et al., 6 Jun 2026). Together these systems shift IE away from purely textual prompting toward structured, inspectable control surfaces.

4. Domain-specific instantiations

The same IE vocabulary appears across heterogeneous technical domains, but each domain chooses its own carrier, operational semantics, and evaluation regime.

Domain	Intent mechanism	Representative paper
Dense retrieval	Natural-language instruction plus pluggable introspector	(Pan et al., 2023)
Intent-Based Networking	KPI-centric intent model, drift detection, corrective policy generation	(Dzeparoska et al., 2024)
AI-agent authorization	Intent certificate, session-scoped narrowing, manifest filtering, payload checks	(Zhu et al., 22 Jun 2026)
OS agents	Structured intents over Persona × Domain × Task × Complexity	(Luo et al., 9 Jun 2026)
Dialogue intent parsing	Neural-Bayesian program with intent mixtures and transitions	(Hong et al., 2024)
Intent-centric software engineering	Desired outcomes, constraints, quality attributes, policies, acceptance evidence	(Cruz, 10 May 2026)
Multimodal safety	Captioning, intent inference, intent-conditioned response refinement	(Na et al., 21 Jul 2025)

In retrieval, the task-specific relevance definition itself is the intent, and instruction templates such as “Retrieve a Wikipedia paragraph to verify this claim” or “Retrieve an argument that counter-argues the following paragraph” provide the control signal (Pan et al., 2023). In networking, intent is represented by target KPI vectors, and “intent drift” is the deviation of operational state from target state, measured through KPI differences, distances, and gradients that inform corrective action (Dzeparoska et al., 2024). In tool-using AI agents, authorization is narrowed by server-side intent certificates that bind request hashes, intent classes, resource bounds, effect bounds, confidence, review mode, and expiry to subsequent tool calls (Zhu et al., 22 Jun 2026).

The software-engineering literature extends IE from user-facing interaction to socio-technical governance. One study characterizes the transition as movement “from code-centric production toward intent-centric human-agent work,” in which engineers increasingly supervise, validate, and govern systems rather than author isolated code artifacts (Cruz, 10 May 2026). Another paper locates IE at level 3 of a cumulative pyramid above prompt engineering and context engineering, arguing that well-contextualized agents still need encoded goals and trade-off hierarchies to act strategically rather than merely efficiently (Vishnyakova, 10 Mar 2026).

Additional variants broaden the field further. ISE makes structured intent the driver of synthetic multi-turn OS-agent trajectories through a 4D intent space and execution-grounded simulation (Luo et al., 9 Jun 2026). DI-Parser treats dialogue intent parsing as probabilistic program learning under data scarcity (Hong et al., 2024). SIA frames multimodal safety as intent-aware reasoning over the joint semantics of images and text (Na et al., 21 Jul 2025). This suggests that IE is less a single method than a family of design and modeling practices that elevate intent to a first-class variable.

5. Empirical evidence and evaluation

The interface-centered study “Towards Intent-based User Interfaces” reports different empirical patterns by task type. In news headline generation, $O$0 participants generated headlines under six conditions and $O$1 expert raters assessed attractiveness, clarity, accuracy, and inoffensiveness; the main finding was that “one-off intent expression via a simple textbox (chatbot-like) sufficed to produce high-quality headlines on average,” while “guidance+selection+post-editing offered additional benefits at modest cost” (Ding, 2024). In cross-domain analogy generation, median helpfulness was $O$2, approximately $O$3 of cases led to substantial reframing, and up to $O$4 of outputs were potentially harmful, indicating that diversity benefits required selection, critique, and guardrails (Ding, 2024). In exploratory visual data analysis with $O$5 analysts, experts produced high-quality but potentially narrow outputs through detailed intents, whereas novices explored broadly with lower-quality outputs and stronger needs for structured guidance (Ding, 2024).

In retrieval, I3 reports average zero-shot nDCG@10 of $O$6 across BEIR, exceeding COCO-DR at $O$7, with per-task highlights including SciFact $O$8, Fever $O$9, Quora $i \in I$0, NQ $i \in I$1, and TREC-COVID $i \in I$2 (Pan et al., 2023). The same paper reports that removing the introspector drops average performance to $i \in I$3, and that with pruning query encoding time decreases from $i \in I$4s to $i \in I$5s while document encoding and retrieval latency remain unchanged because the document encoder is frozen (Pan et al., 2023). In UI understanding, UI-JEPA establishes the IIW dataset with 1.7K videos across 219 intent categories and the IIT dataset with 914 videos across 10 categories, and reports that, averaged across two datasets, it outperforms GPT-4 Turbo and Claude 3.5 Sonnet by $i \in I$6 and $i \in I$7 respectively, while achieving a $i \in I$8 reduction in computational cost and a $i \in I$9 improvement in latency on IIW (Fu et al., 2024).

Execution-grounded and perception-grounded IE also reports concrete gains. ISE constructs roughly 50000 structured intents, retains 43956 unique intents after deduplication, achieves a Vendi Score of 61.57 over the full intent pool, and generates 23132 complete OS-agent trajectories averaging 8.12 user turns and 68.24 total dialogue turns (Luo et al., 9 Jun 2026). Fine-tuning on ISETrace improves ClawEval pass@1 from 19.3 to 37.7 with Qwen3-8B, while truncating the data to a single user turn reduces pass@1 to 28.1 (Luo et al., 9 Jun 2026). DI-Parser, on 952,749 financial customer service dialogues with only 1,000 dialogues annotated via AMT, reports 79.6% accuracy versus 77.9% for a Transformer baseline (Hong et al., 2024).

Multimodal and safety-oriented results point in the same direction. IEA reports stage-wise improvement in image-edit pixel distance from 0.134 at Stage 1 to 0.103 at Stage 3, and in Image-Summary ROUGE-L from 0.222 to 0.258, with Friedman tests yielding $i=(G,C,P)$0 in the user study (Zhu et al., 6 Jun 2026). SIA, a training-free intent-aware safety method for VLMs, reports substantial gains on SIUO, HoliSafe, and MM-SafetyBench; for Gemma3-IT-4B on SIUO, safety/effectiveness moves from 28.14 / 93.41 in the base model to 62.28 / 97.60 with SIA, while MMStar average accuracy drops from 36.88 for Image+Query to 33.41 for Caption+Query+Intent (Na et al., 21 Jul 2025). These results do not collapse to a single metric, but they collectively support the claim that explicit intent modeling can improve task quality, safety, or controllability when the evaluation target is chosen to match the domain.

6. Governance, failure modes, and future directions

A major theme in IE is that control cannot be delegated entirely to the model. IGAC, for tool-using agents, enforces the invariant that user intent may only reduce authority relative to static policy, operationalized through intent certificates, session-scoped policy narrowing, intent-aware manifest filtering, and intent–tool–payload consistency checks (Zhu et al., 22 Jun 2026). In enterprise multi-agent architecture, IE is explicitly tied to governance, provenance, attestation, privilege attenuation, and business-level trade-off hierarchies; it occupies the level above context engineering in a cumulative maturity model, and its function is strategic rather than merely informational (Vishnyakova, 10 Mar 2026).

The failure modes are correspondingly diverse. In HCI settings, over-fixation on precise specifications can narrow exploration, vague intents can yield low-quality output, and unfiltered creative outputs can cause harm (Ding, 2024). In retrieval, misleading or low-quality instructions degrade performance, incorrect intent specification leads to inappropriate ranking, and very fine-grained or compositional intents may exceed what a single introspected embedding can capture (Pan et al., 2023). In on-device UI understanding, zero-shot generalization remains weaker than few-shot performance, especially on IIT, and fine-grained text-heavy tasks depend on OCR quality (Fu et al., 2024). In multimodal safety, prompt sensitivity, caption quality, and domain shift can produce both over-refusal and missed harms (Na et al., 21 Jul 2025). Intent Signal Theory adds a stronger claim: high structural quality does not guarantee fidelity to private intent, and holistic quality metrics can obscure this structural–fidelity split (Peng, 24 May 2026).

Several papers outline convergent future directions. The IBN assurance work points to proactive assurance, digital twins, and tool-augmented LLMs for quantitative planning with guarantees (Dzeparoska et al., 2024). The software-engineering analysis emphasizes “bounded autonomy,” evidence gates, provenance logging, policy-as-code, and longitudinal evaluation of technical debt, review quality, and accountability (Cruz, 10 May 2026). The IUI paper identifies improved suggestions that elicit broad exploratory intents from experts, structured iteration guidance for novices, larger mixed-methods evaluations, and reusable components such as intent canvases, provenance mechanisms, and suggestion widgets as next steps (Ding, 2024).

Across these strands, IE appears as a control-and-evaluation layer for systems that act under underspecification, ambiguity, and delegation. This suggests a unifying interpretation: IE is the discipline of making intent explicit enough to be operational, inspectable, and governable, while preserving the distinction between what is sought, how it is encoded, what the system does, and how fidelity is assessed.