Progressive Ideation using an Agentic AI Framework for Human-AI Co-Creation

TL;DR

This paper tackles the challenge of generating truly novel design ideas by criticizing the prevalent single-shot AI ideation paradigm and introducing MIDAS, a distributed, agentic framework that emulates human meta cognitive ideation. MIDAS replaces a lone LLM with a coordinated team of specialized agents operating in a continuous CG/CA pipeline under a Participatory, Active, Collaborative (PAC) partnership with the designer. It formalizes problem representation (AI3C AOC PFIC), applies a fourfold NDFR evaluation (Novelty, Diversity, Feasibility, Relevance), and grounds ideation against real-world prior art through Librarian and Challenger agents. Demonstrations across six design problems with six novice designers show MIDAS producing highly diverse and globally novel concepts within 20 minutes, suggesting significant practical benefits for design practice and pedagogy while highlighting avenues for future research in multi-modal and larger-scale validation.

Abstract

The generation of truly novel and diverse ideas is important for contemporary engineering design, yet it remains a significant cognitive challenge for novice designers. Current 'single-spurt' AI systems exacerbate this challenge by producing a high volume of semantically clustered ideas. We propose MIDAS (Meta-cognitive Ideation through Distributed Agentic AI System), a novel framework that replaces the single-AI paradigm with a distributed 'team' of specialized AI agents designed to emulate the human meta-cognitive ideation workflow. This agentic system progressively refines ideas and assesses each one for both global novelty (against existing solutions) and local novelty (against previously generated ideas). MIDAS, therefore, demonstrates a viable and progressive paradigm for true human-AI co-creation, elevating the human designer from a passive filterer to a participatory, active, collaborative partner.

Figures (10)

• Figure 1: Sankey Map illustrating the interconnections between modality terms (e.g., AI, Generative AI, LLM), interaction terms (e.g., Collaborative, Augmented, Symbiotic), and design terms (e.g., Ideation, Co-Design, Innovation) in AI-assisted creative processes.
• Figure 2: Networked, non-linear ideation via “ideas-from-ideas.” Unlike the prevailing linear model, where a single problem $P_1$ yields a one-way list of ideas $\{I_1,\dots\}$, this diagram depicts ideation as a reusable, cross-problem network. Problems ($P$) are shown on the left and idea nodes ($I$) on the right, with edges indicating applicability. Crucially, some ideas generated for one problem are also viable for other problems (e.g., $I_4$ links $P_1$ and $P_2$; $I_5$ links $P_1$ and $P_3$), enabling conceptual blending and analogical transfer. This “re-association” creates pathways where ideas discovered under $P_2$ (e.g., $I_6$--$I_8$) can be adapted back to address $P_1$ (e.g., $I_7$), and similarly across other problem–idea constellations (e.g., $I_{13}$ links $P_2$ and $P_3$, while $I_{11}$ links $P_1$ and $P_3$). Node colors encode whether an idea is unique to a single problem or shared across problems, highlighting bridges that enable non-linear creative leaps.
• Figure 3: A comparative visualization of semantic clustering for idea generation methodologies for the problem statement PS1 (Waste Segregation) and PS2 (Footwear Disinfection). (a), (c) 'Single-spurt' generation results in high-density clusters of similar ideas. (b), (d) 'Progressive provocation' forces the generation of semantically distant and diverse ideas (blue dots). The transition from high-density clusters in the baseline to diverse micro-clusters in the progressive sets is evident across both domains.
• Figure 4: A comparative visualization of semantic clustering for idea generation methodologies for the problem statement PS3 (Dish Cleaning) and PS4 (Queuing Comfort). (a), (c) 'Single-spurt' generation results in high-density clusters of similar ideas. (b), (d) 'Progressive provocation' forces the generation of semantically distant and diverse ideas (blue dots). The transition from high-density clusters in the baseline to diverse micro-clusters in the progressive sets is evident across both domains.
• Figure 5: A comparative visualization of semantic clustering for idea generation methodologies for the problem statement PS5 (Bird Feeding) and PS6 (Umbrella Storage). (a), (c) 'Single-spurt' generation results in high-density clusters of similar ideas. (b), (d) 'Progressive provocation' forces the generation of semantically distant and diverse ideas (blue dots). The transition from high-density clusters in the baseline to diverse micro-clusters in the progressive sets is evident across both domains.