'The End of the Static Screen: Architecting Intent-Driven UX - Gus Iwanaga, commercetools'

Building proper generative UX/UI requires moving beyond static software experiences using AI-driven agentic orchestration, but doing so responsibly requires ...

By Sean Weldon

The End of the Static Screen: Architecting Intent-Driven UX

Abstract

For four decades, software has shipped fixed interfaces to which users must adapt, accruing cognitive load as organizations adopt multiple Software-as-a-Service (SaaS) applications with divergent mental models. This synthesis examines an industry approach to constructing generative UX/UI - interfaces assembled at runtime by an agentic orchestration system in response to user intent - drawn from commercetools' implementation experience. The methodology combines prototype evidence, protocol classification, and organizational analysis. Findings indicate that unconstrained large language model (LLM) generation produces non-deterministic, unshippable outputs, and that a declarative UI protocol approach, mediated by a curated component catalog and atomic design hierarchy, offers a viable middle ground between rigid controlled interfaces and unbounded generation. Practical implications include a redefinition of design and product roles around schema authorship and catalog curation rather than direct pixel manipulation.

1. Introduction

Conventional enterprise software delivery assumes an interface authored in advance and consumed uniformly across a user base. This model has persisted for approximately forty years, producing an asymmetry succinctly captured in the observation that "we're still adapt to the software that we ship, not the other way around." The release of GPT in November 2022 provided the first widely visible demonstration of dynamic personalization, yet the majority of enterprise software has remained architecturally static in the years since.

This persistence is compounded by SaaS proliferation. Organizations typically operate several applications concurrently - a customer relationship management (CRM) tool, a table-dense operational system, a polished consumer-facing product - each imposing distinct navigational conventions and terminology. Comparative review of such tools indicates that information overload and onboarding cost persist regardless of visual polish, suggesting the underlying problem is architectural rather than aesthetic.

This raises the central question examined here: what foundational shifts become possible if software interaction is mediated dynamically by AI at runtime, rather than fixed at build time? Generative UX/UI is defined for this analysis as an interface composition process in which an agentic orchestrator interprets user intent, retrieves relevant context and tools, and assembles a corresponding interface on demand. The analysis proceeds through prototype evidence documenting failure modes of unconstrained generation, a classification of emerging UI protocols, and three implementation challenges - information architecture, catalog curation, and organizational restructuring.

2. Background and Related Work

Three bodies of practice inform this approach. First, atomic design methodology, a five-stage framework for constructing interface design systems as hierarchical compositions of primitives (atoms through templates), supplies the structural vocabulary used to decompose generated interfaces into layout, slots, sub-slots, and components. Second, the Model Context Protocol (MCP) and adjacent tool-invocation standards provide the mechanism by which an orchestrating agent discovers and invokes first-party services, third-party services, and subordinate agents to gather task-relevant context. Third, an emerging class of declarative UI protocols - including HTMX (Google), JSON Render (Vercel), and OpenUI (Thesis) - defines schema-mediated contracts, often specified via constructs such as Zod schemas, through which an agent selects and configures interface components without generating markup directly.

Collectively, these establish the intellectual premise underlying this analysis: generative UI is fundamentally a protocol and governance problem rather than a pure model-capability problem.

3. Core Analysis

3.1 Prototype Evidence and the Non-Determinism Failure Mode

Early prototypes generated interfaces on demand for natural-language queries such as "create a sales report for Q1." Successive invocations of semantically identical queries produced divergent outputs - the system alternated between interpreting the period as "Q1" and as "January-March," and rearranged component layout between turns. This output was assessed as unshippable, since inconsistent component arrangement undermines the predictability required in production B2B software. The current architecture addresses this through an orchestrator that extracts user intent, locates applicable tools (first-party, third-party, agents, or MCP servers), and passes assembled context to a dedicated UX agent responsible for rendering. A subsequent demonstration involving campaign-planning queries showed markedly improved consistency and aesthetic coherence, including an approval workflow that pushes AI-generated UI to a pre-production environment for review before release.

3.2 Classification of UI Protocol Approaches

Three architectural approaches to generative UI rendering are identified. The controlled approach constrains the agent to selecting a single pre-built, opinionated component rendered as-is - exemplified by consumer-facing implementations such as a restaurant-finder demonstration, suitable where interaction scope is narrow. The open-ended approach grants the LLM full generative autonomy over interface output, exemplified by an organizational-chart generation demonstration; this approach is characterized as risky because neither output nor downstream outcome can be reliably controlled. The declarative approach occupies a middle position: protocols such as HTMX, JSON Render, and OpenUI allow the orchestrator to map retrieved data onto a defined component catalog via a UI specification, which is then rendered as native components (e.g., React) compliant with an established design system. This third approach was selected for B2B deployment because enterprise customers require consistency - identical copy, stable date ranges, predictable layout - properties incompatible with unconstrained LLM creative control.

3.3 Information Architecture and Component Placement

A central implementation question concerns what governs component arrangement once an agent selects from a catalog. Ungoverned placement produces confusing output, as prototype evidence demonstrated. The response was adoption of atomic design methodology to structure the interface hierarchically: page/layout, slots, sub-slots, and eligible component categories. This hierarchy is then inverted for orchestration purposes, so that selected components map back upward through sub-slots and slots to templates. The UX agent is trained against a curated catalog of layout templates representing "what good looks like," constraining generative freedom to a bounded, pre-validated design space.

3.4 Design System Curation as Contractual Infrastructure

The component catalog functions as "the contract between the agent and the UI," such that every property definition carries functional consequence. This necessitates continuous testing across candidate UI protocols (referenced as ATUI, JSON Render, and OpenUI) to refine reliability, alongside careful curation of layout components' own attributes at each hierarchical level. This curation effort is characterized as the differentiator between a superficial demonstration and a viable production experience - the underlying protocol capability is necessary but insufficient without disciplined catalog governance.

4. Technical Insights

5. Discussion

These findings suggest that the maturation of generative UI depends less on advances in raw model capability and more on the surrounding governance infrastructure - protocol design, catalog curation, and information architecture. This reframes generative UI from a model-selection problem to a systems-engineering problem, consistent with broader trends toward structured tool use and schema-constrained generation observed across agentic AI system design generally.

The organizational consequence documented is substantial: design teams no longer author pixels directly, since AI generation now dictates much of the visual output. Work shifts toward schema definition, catalog curation, rule-setting, and synthetic data generation for query-to-component mapping. This transition poses a documented adjustment challenge, particularly for non-technical product managers and designers accustomed to direct visual authorship. The proposed Three P's framework - People, Product, Process - is offered as an organizational response, emphasizing that lightweight process changes must accompany shifts in people's roles and product architecture; however, the source material does not provide detailed metrics on adoption timelines or measured productivity outcomes, representing an open area for further empirical investigation.

6. Conclusion

This analysis documents a transition from static, user-adapted software toward intent-driven interfaces assembled dynamically by orchestrated AI systems. The central contribution is the identification of the declarative UI protocol as a viable middle ground between rigid, controlled components and unconstrained generative autonomy, validated through prototype evidence showing that unconstrained approaches produce non-deterministic, unshippable output. Practically, organizations pursuing generative UX should prioritize catalog curation and hierarchical information architecture before protocol selection, and should anticipate substantial role redefinition for design and product functions. Future work should examine measurable reliability benchmarks across competing declarative protocols and longitudinal organizational adoption outcomes.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub