• 26 Dec, 2025 *

Abstract

This paper develops a research position at the intersection of urban spatial science, planning theory, and agentic artificial intelligence. It examines a novel paradigm for complex institutional software systems in which user interfaces are not designed ex ante, but instead converge through an agentic, perception‑driven feedback loop. Drawing on traditions within CASA concerned with cities as computational, representational, and procedural systems, the paper analyses a prototype in which a large language model (LLM) acts simultaneously as planner, spatial reasoner, and code‑level implementer. Using browser‑level perception (via Playwright‑based Model Context Protocols) and constrained code mutation, the system iteratively reshapes its own interface in response to rendered output, expert judgement, and backend procedural requirements. The contribution is not a UI design method, but a computational account of interface formation as an emergent property of institutional reasoning systems, with particular relevance to spatial planning and governance.

1. Introduction

Research at the Centre for Advanced Spatial Analysis (CASA) has long treated cities, institutions, and planning systems as computational artefacts: shaped by representations, rules, feedback loops, and bounded rationality. Within this tradition, interfaces are not neutral surfaces but critical mediators between data, models, and human judgement. Yet most digital planning tools continue to rely on interface designs specified in advance, assuming that usability, transparency, and legitimacy can be engineered top‑down.

In high‑stakes institutional domains such as spatial planning, this assumption is increasingly untenable. Planning practice requires interfaces that simultaneously satisfy aesthetic dignity, operational usability, and deep procedural transparency, while remaining adaptable to context‑specific judgement and political framing. This paper argues that such interfaces cannot be fully specified a priori and instead must be allowed to emerge through controlled interaction between reasoning agents, rendered artefacts, and expert oversight.

We present an approach in which a model is granted both interpretive and generative authority over a live codebase, forming a closed perceptual–action loop grounded in the rendered interface itself. Rather than producing an optimised product UI, the system gives rise to an evolving administrative instrument whose structure stabilises through convergence rather than prescription. This framing aligns with CASA’s broader interest in dynamic systems, feedback‑driven urban processes, and the computational constitution of institutional behaviour.

2. Conceptual Foundations

2.1 Agentic Programming

Recent work in agentic programming has focused on autonomous or semi‑autonomous systems capable of multi‑step reasoning, tool use, and environment interaction. Most existing approaches, however, constrain agents to symbolic representations of state (e.g. abstract syntax trees, test results, or logs). The present work departs from this by insisting that the agent’s primary epistemic input is the rendered user interface, accessed through browser automation and screenshot‑based perception. This shifts the agent’s locus of judgement from intent to consequence.

2.2 Human–Computer Interaction

Within HCI, progressive disclosure and usability heuristics are often treated as design objectives to be balanced through user testing and iteration. This paper reframes these concerns as simultaneous constraints evaluated continuously rather than sequentially. Beauty, usability, and depth are not trade‑offs but co‑present requirements, reflecting the expectations of professional administrative users rather than consumer audiences. Evaluation occurs phenomenologically, through expert judgement of the rendered artefact, rather than through abstract usability metrics.

2.3 Planning Theory and Institutional Rationality

Planning theory provides a critical lens for understanding why conventional UI optimisation fails in institutional settings. Planning judgement is procedural, discursive, and legally conditioned; it depends on the visible ordering of evidence, reasons, and authority across time rather than at a single moment of interaction. Interfaces that obscure this ordering, even if functionally complete, undermine trust and institutional legitimacy.

From a CASA perspective, this can be understood as a failure of representational alignment between computational systems and institutional rationality. The system described here encodes planning concepts explicitly in its backend while relying on latent model priors—shaped by exposure to administrative artefacts—to propose interface arrangements that feel procedurally plausible to practitioners. Interface convergence is thus treated as a spatial–temporal alignment problem between representations, rather than a question of visual optimisation.

2.4 Live System Introspection

Live system introspection refers to the ability of a system to observe and reason about its own operational state in real time. In this work, introspection is extended to the interface layer itself: the system perceives how its reasoning is presented, evaluates that presentation against domain‑specific expectations, and mutates its own code accordingly. This creates a feedback loop analogous to cybernetic control systems, in which stability emerges through repeated, bounded adjustments.

3. System Architecture

The proposed system comprises four tightly coupled components:

1. Perceptual Layer: Browser‑level observation using Playwright MCP, providing screenshots and DOM state as ground truth.
2. Reasoning Agent: A large language model instantiated with planner‑aligned priors, capable of evaluating procedural legibility and interface coherence.
3. Mutation Mechanism: Constrained code modification privileges, limiting the scope and radius of changes to encourage convergence.
4. Human Oversight Loop: Expert human judgement, expressed through rapid acceptance or rejection of changes (e.g. via version control), acting as a stabilising oracle.

Crucially, no static UI specification is privileged over this loop. Specifications are inferred post hoc from stable interaction patterns that survive repeated perturbation.

4. Evaluation and Convergence

Evaluation in this paradigm is explicitly multi‑objective. Each iteration is judged simultaneously against:

• Aesthetic Dignity: Does the interface appear calm, credible, and appropriate to an institutional context?
• Operational Usability: Are primary actions discoverable and friction‑free for expert users?
• Progressive Depth: Can all relevant data and AI‑assisted reasoning be accessed without overwhelming the initial view?

Rather than formal metrics, convergence is assessed phenomenologically by expert practitioners. Reversibility is central: changes are inexpensive to discard, encouraging aggressive exploration without architectural commitment. Over time, oscillation dampens as the interface approaches a stable attractor satisfying all constraints.

5. Discussion

This approach challenges several assumptions in both software engineering and HCI. First, it rejects the notion that interface design must precede implementation. Second, it demonstrates that model priors—often treated as a liability—can be productively harnessed when aligned with domain practice. Finally, it suggests that institutional interfaces may be better understood as living artefacts whose legitimacy arises from procedural resonance rather than visual optimisation.

The method is not without risks. Unconstrained agents may drift toward novelty or aesthetic excess, while overly restrictive mutation bounds can stall convergence. However, these risks are manageable through careful control of perceptual inputs and human veto mechanisms.

6. Conclusion

This paper has outlined a convergent, agent‑driven paradigm for interface formation situated within the intellectual traditions of urban spatial science and computational planning. By collapsing the boundaries between planning, coding, and evaluation, and grounding judgement in rendered artefacts rather than abstract specifications, the approach treats interfaces as emergent properties of institutional reasoning systems.

For CASA, the contribution lies in reframing interface design as a dynamic, feedback‑driven process analogous to urban system evolution: stabilised through repeated interaction rather than imposed structure. While demonstrated in the context of spatial planning, the framework has broader relevance for research into digital governance, institutional modelling, and the computational representation of judgement. Future work will explore formal characterisations of convergence dynamics, links to complexity theory and cybernetics, and empirical comparison with conventional digital planning workflows.