Architectural researchers have introduced a conversational agentic artificial intelligence framework that transforms early-stage building design from rigid commands into an iterative, multi-party dialogue. Published in Architectural Intelligence by Lok Hang Cheung, Likai Wang, and Dongxue Lei, the study addresses critical bottlenecks in computational architecture, where generative software often operates in isolation from site context and environmental performance requirements. By integrating an architecture-centric multimodal agent system directly into WeChat, the team established an accessible collaborative medium connecting undergraduate architecture students, design instructors, and an artificial intelligence partner. Rather than functioning as an automated drafting tool or basic image generator, the system actively participates in early massing evaluation, performative optimization, and stylistic testing.
The development of computer-aided architectural design has progressed substantially since the nineteen-sixties, yet contemporary generative artificial intelligence workflows frequently reproduce reductive input-output interactions. Previous investigations highlighted that architects must switch continuously between disparate complex platforms, facing steep learning curves and mastering technical deep-learning scripts merely to maintain iterative feedback loops. Furthermore, standalone generative visual models frequently produce site-agnostic configurations that sever architectural language from environmental constraints and existing urban fabric. To overcome these deficiencies, the researchers operationalized theoretical propositions that conceptualize design as an open-ended negotiation, reframing computational tools from passive drawing instruments into reflective, mutual-learning design collaborators.
This investigative conceptual framework directly revives classic architectural cybernetics and foundational design theories, notably Gordon Pask’s Conversation Theory and Donald Schön’s formulation of design as a reflective conversation. Pask theorized computational collaboration as an iterative exchange wherein human participants and computational entities achieve shared comprehension through continuous mutual learning, while Schön characterized design as a reflective transaction with situational materials. The authors also draw upon Nicholas Negroponte’s vision from the Architecture Machine Group, which sought intelligent computational partners capable of deliberating over architectural constraints rather than acting as fast automated draftsmen. Contemporary architectural discourse has similarly called for generative artificial intelligence interactions characterized by negotiation and deliberation, prioritizing creative novelty, context awareness, and collaborative meaning-making over prediction accuracy. By situating multimodal models within active group dialogue, the researchers established an experimental environment where design intentions, environmental simulations, and aesthetic variations undergo sustained, multifaceted refinement.
To implement this framework, the researchers developed an architecture-focused agentic multimodal system built upon the LinkAI platform and deployed through the ChatGPT-on-WeChat interface. The computational architecture combines OpenAI’s GPT-4 multimodal models with the DALL-E 3 image generator, incorporating retrieval-augmented generation grounded in a specialized architectural knowledge base. This repository contains project briefs, performative design literature, and the operational manual for EvoMass, an environmental massing optimization tool. System interactions are coordinated through three functional components: an initial multimodal model serves as a classifier inspecting submissions to identify whether visual data represents an abstract massing model or a developed rendering. When an early massing model is detected, the platform triggers a structured workflow executing a Google search to retrieve three precedent studies matching geometric features, synthesizing contextual advice across three distinct directions before producing a reference illustration.
When designers submit developed renderings, the system activates a second workflow that bypasses external precedent retrieval while delivering qualitative evaluations and targeted stylistic imagery. If inputs contain unconventional queries, mixed media, or general conceptual prompts, requests route to a third multimodal model operating as an autonomous agent that dynamically invokes web search or image synthesis tools. To connect algorithmic performance with architectural expression, the authors established a two-stage methodology linking massing optimization with downstream stylistic exploration. Abstract volumetric models generated via EvoMass prioritize solar exposure and environmental performance, but lack facade articulation and material language. Designers export these optimized massing geometry profiles into the WeChat dialogue, where the agent analyzes spatial attributes, suggests compositional strategies, and formulates descriptive prompts. Architects subsequently transfer these prompts into Stable Diffusion paired with ControlNet models to execute controlled stylistic variations across structural depth and edge parameters.
The empirical viability of this hybrid workflow was verified through a design workshop at Xi’an Jiaotong-Liverpool University involving three undergraduate architecture students and studio instructors. The workshop focused on renovating a mixed-use building cluster in the historic Gusu Old Town in Suzhou, China, a dense setting marked by inadequate neighbourhood public spaces and dim side alleys shadowed by overgrown foliage and encroaching structures. Traditional renewal efforts in the region have frequently prioritized superficial facade repainting while neglecting structural typology and environmental living quality. Participants were tasked with generating proposals balancing urban daylight performance and community spaces with the preservation of traditional Chinese architectural characteristics. Students and instructors joined a shared WeChat group alongside the artificial intelligence bot, allowing participants to converse with the system, monitor each other’s developmental iterations, and draw inspiration from peer interactions within an everyday mobile communication interface.
Across the workshop, the researchers documented three distinct conversational design models, beginning with a performance-based approach adopted by Student 1. Prioritizing neighborhood daylight enhancement, Student 1 used EvoMass to evaluate building orientation, submitting preliminary massing outcomes to the bot for developmental guidance. Because the submission combined analytical diagrams with narrative inquiries, the query activated the autonomous third model, which interpreted simulation data and recommended stepped setbacks alongside EvoMass’s rotation module to diminish shadowing on adjacent residences. Unaware of the rotation feature previously, the student manually adjusted road-facing modules and shifted from additive to subtractive volume generation modes. Subsequent massing submissions triggered the precedent-retrieval workflow, producing relevant case studies that guided further refinements. When algorithmic responses became overly generic, studio tutors intervened by posing clarifying technical questions, demonstrating how instructors can effectively scaffold student interactions with computational agents.
A second participant demonstrated an evaluation-based approach, prioritizing usable floor area, sunlight hours, and diagonal circulation arteries connecting internal community spaces with surrounding street networks. Seeking to compare alternative massing configurations, Student 2 submitted comparative screenshots; however, because initial perspectives were captured from low pedestrian eye-levels, the visual classifier struggled to separate proposed structures from contextual background buildings. The student resolved this ambiguity by adding horizontal floor slab lines onto the massing model and instructing the bot to evaluate the annotated structure. This interaction triggered the precedent-driven workflow, generating architectural analyses that Student 2 translated into detailed prompt directives for Stable Diffusion style testing. As the project progressed, the student treated the bot as an analytical team member, requesting it to formulate parallel prompt strings automatically and submitting multiple candidate renders to explore parallel variations balancing density with historic context.
The third participant exhibited a designerly-based methodology, possessing an established conceptual vision centered on open public streetscapes and extensive previous proficiency with Stable Diffusion. Initially indifferent to the conversational chatbot, Student 3 bypassed the agent during early massing phases, relying on autonomous parametric modeling and personal prompt craft. However, after witnessing the bot’s comparative analytical capacities in peer exchanges, Student 3 began submitting rendered facades to interrogate its evaluative criteria. When the artificial intelligence recommended a preferred scheme from five options, the student critically challenged the assessment, asking how the bot balanced open street life against the disruption of traditional residential patterns. Engaged almost exclusively with the flexible autonomous model, Student 3 demanded explicit architectural methods to reconcile these community tensions, translating conversational insights into targeted facade prompts and testing synthetic green space integrations between public roadways and building perimeters.
A quantitative analysis of workshop communications revealed significant operational patterns regarding how designers interact with conversational multimodal agents. Defining usable interactions as exchanges completed without technical failure, the investigators recorded that multimodal visual inputs constituted over sixty percent of total useful interactions, confirming that visual processing is indispensable for architectural deliberation. Conversely, multimodal visual outputs generated directly by the bot accounted for only one-third of interactions, as participants preferred conducting graphic iterations within dedicated external software like Stable Diffusion. Furthermore, nearly eighty percent of conversational exchanges were processed by the autonomous third model rather than predetermined workflows, demonstrating that fluid language models possess superior adaptability across dynamic design scenarios. Collaborative contributions from peers and tutors occurred in roughly one-third of total interactions, underscoring the pedagogical utility of shared group environments, though researchers noted the lack of a control group prevents definitive comparative conclusions.
Post-workshop semi-structured interviews and standardized questionnaires provided qualitative nuance, revealing both strong enthusiasm and notable friction. All three participants reported that the conversational interface alleviated decision-making fatigue by offering objective secondary evaluations, yet Students 2 and 3 cautioned that algorithmic suggestions occasionally lacked logical consistency, underscoring the necessity of integrating explainable artificial intelligence and chain-of-thought verification mechanisms. While Students 1 and 2 valued the WeChat group for mutual peer learning, Student 3 experienced cognitive overload from simultaneous multimodal threads, advocating for dual private and public chat architectures. Participants also noted domain comprehension deficiencies, such as the bot proposing stereotypical garden motifs or irrelevant safety codes during massing queries. Questionnaire results showed the framework achieved slightly higher effectiveness than traditional methods with equivalent learning curves. Future work will focus on integrating image synthesis engines directly, supporting multi-image reasoning, and embedding domain knowledge graphs.
Subject of Research: Conversational agentic multimodal artificial intelligence for early-stage architectural design ideation and performative exploration
Article Title: Conversational, agentic AI-enhanced architectural design process: three approaches to multimodal AI-enhanced early-stage performative design exploration
Article References: Cheung, L. H., Wang, L., & Lei, D. (2025). Conversational, agentic AI-enhanced architectural design process: three approaches to multimodal AI-enhanced early-stage performative design exploration. Architectural Intelligence, 4(1), Article 10. https://doi.org/10.1007/s44223-025-00092-5
Image Credits: AI Generated
DOI: 10.1007/s44223-025-00092-5
Keywords: Architectural Design, Multimodal Artificial Intelligence, Agentic AI, Conversational Design Framework, Performative Optimisation, EvoMass, ControlNet, Stable Diffusion, Building Massing, WeChat Integration, Design Ideation, Human-AI Collaboration, Generative AI, Computer-Aided Architectural Design, Suzhou Old Town Renovation, Urban Daylight Performance, Large Language Models, Explainable AI
News Source: Denise Maddox. (November 11, 2026). Multimodal Agentic AI Framework Enhances Early Architectural Design Exploration in Collaborative Messaging. Scienmag.
