Beyond Brainstorming: AI Teams That Build on Ideas

Author: Denis Avetisyan

A new agentic AI framework moves past simple idea generation, fostering a collaborative process that unlocks demonstrably more novel and diverse concepts.

The MIDAS framework establishes a collaborative environment where human ideation is amplified by a distributed network of thirteen specialized AI agents-orchestrated by a central ‘Professor’ agent and linked via persistent ‘Vaults’-across six operational phases, effectively externalizing and accelerating the creative process.

This paper introduces MIDAS, an agentic AI system leveraging a distributed team of specialized agents for progressive ideation and human-AI co-creation.

While truly novel idea generation remains a key challenge in contemporary design, current AI systems often produce semantically clustered results, hindering creative exploration. This paper introduces ‘Progressive Ideation using an Agentic AI Framework for Human-AI Co-Creation’ and details MIDAS, a novel framework employing a distributed ‘team’ of specialized AI agents to emulate human meta-cognitive ideation. MIDAS progressively refines concepts, assessing both global and local novelty, demonstrably improving the diversity and originality of generated ideas. Could this agentic approach represent a paradigm shift towards genuine human-AI collaborative creativity, elevating designers from passive reviewers to active co-creators?

Deconstructing Ideation: Beyond the Single Spark

Conventional artificial intelligence approaches to ideation, frequently termed ‘Single-Spurt Generation’, prioritize quantity over quality, often yielding a large number of concepts that lack substantial development or nuanced detail. These systems typically operate by generating ideas in a single, rapid burst, failing to build upon initial concepts or explore variations effectively. While capable of producing a broad range of suggestions, this method often results in a collection of superficial ideas, hindering the potential for truly innovative breakthroughs. The emphasis on sheer volume overshadows the iterative process crucial to human creativity, where concepts are refined, combined, and adapted over time – a capability currently limited in many AI-driven ideation platforms.

Human creative processes rarely manifest as a single burst of inspiration; instead, innovation typically unfolds through iterative cycles of conceptual development. Existing ideas are rarely discarded outright, but rather refined, combined, and adapted, building upon a foundation of prior knowledge and experience. This progressive approach, where each iteration informs the next, allows for deeper exploration of a concept space and ultimately yields more nuanced and robust solutions. Current artificial intelligence systems often lack this crucial capacity for sustained, iterative refinement, frequently generating ideas in isolation without the ability to learn from, or build upon, previous outputs – a limitation that hinders their potential for truly groundbreaking innovation and necessitates a shift towards methods mirroring the naturally iterative nature of human thought.

Current artificial intelligence often approaches ideation as a single, exhaustive burst, generating numerous concepts without the crucial element of iterative refinement that characterizes human creativity. This contrasts sharply with truly effective innovation, which necessitates building upon existing knowledge and continuously improving ideas. Recent advancements, such as the MIDAS system, demonstrate a departure from this ‘single-spurt’ methodology; analysis using the DBSCAN algorithm reveals that MIDAS generates an idea pool so semantically diverse it is classified as ‘noise’ – a state signifying a radical break from the dense clusters of similar concepts typically produced by conventional AI. This ‘noise’ isn’t a flaw, but rather an indicator of expansive exploration and a potential pathway towards genuinely novel solutions, highlighting the importance of frameworks that prioritize continuous refinement and knowledge leveraging.

Semantic clustering reveals that while a single-spurt approach to idea generation yields concentrated, similar concepts, progressive provocation effectively diversifies ideas across problem statements, transitioning from dense clusters to sparse, micro-clusters.

MIDAS: Orchestrating a Distributed Creative Network

The MIDAS framework employs a distributed ideation process facilitated by a network of specialized agents. These agents are not general-purpose AI; rather, each is designed to fulfill a specific role within the creative workflow, such as problem decomposition, concept generation, or solution evaluation. This specialization allows for a modular and scalable approach to ideation, enabling parallel processing of tasks and focused expertise at each stage. Communication and data exchange between agents are managed by the framework, coordinating their activities to achieve a unified outcome. The architecture supports various agent types, including those utilizing techniques like $NLP$ , knowledge graphs, and computational creativity algorithms, tailored to their designated function.

The MIDAS framework’s core design model, ‘PAC’, prioritizes a specific interaction dynamic between human designers and AI agents. ‘Participatory’ interaction ensures designers retain creative control and provide direction, rather than passively receiving suggestions. ‘Active’ interaction involves continuous feedback loops and iterative refinement of ideas between humans and AI. Finally, ‘Collaborative’ interaction establishes a shared workspace and common understanding, allowing both designers and agents to build upon each other’s contributions, ultimately fostering a synergistic ideation process.

The MIDAS framework incorporates ‘AI3C’ – a structured problem definition process designed to establish a shared understanding of the ideation challenge prior to solution generation. This process is completed within a standardized 20-minute session and focuses on iteratively refining the problem statement through AI-assisted analysis and human validation. Comparative analysis indicates that utilizing AI3C results in a measurable increase in ideation efficiency when contrasted with traditional brainstorming methods, which often lack a formalized initial problem scoping phase and can suffer from ambiguity or divergent interpretations of the core challenge.

The MIDAS framework facilitates iterative design by enabling communication and knowledge sharing between Generator and Evaluator agents, a persistent Vault, and a human Designer.

Agents of Creation: Dissecting the Creative Process

The ‘Forge’ agent functions as the primary driver of initial concept generation within the system. It achieves this not through a single process, but by dynamically coordinating a network of specialized sub-agents. Specifically, the ‘Explorer’ sub-agent is tasked with identifying and proposing entirely novel concepts, often diverging from established patterns. Complementing this, the ‘Formulator’ sub-agent focuses on translating these abstract ideas into practical, implementable solutions, assessing feasibility and outlining concrete steps for development. This division of labor allows the ‘Forge’ agent to simultaneously pursue both radical innovation and pragmatic application, maximizing the breadth and utility of the generated concepts.

The ‘Mint’ agent functions by systematically dissecting established concepts into their fundamental components, enabling the generation of derivative ideas through recombination and modification. This decomposition process identifies core principles and reusable elements, providing a basis for novel variations. Complementing this, the ‘Scribe’ agent prepares the initial problem statement and any related data into a standardized, computationally-accessible format. This structured representation facilitates efficient processing by subsequent agents within the network and ensures consistency in the evaluation of potential solutions. Both agents operate in tandem during the initial phase of idea generation, transforming existing knowledge and raw problem definitions into actionable inputs.

The evaluation phase utilizes a network of agents, specifically ‘Challenger’ and ‘Sentinel’, to refine an initial set of generated concepts. ‘Challenger’ assesses the novelty of each idea, identifying genuinely new approaches, while ‘Sentinel’ verifies alignment with the original problem statement, discarding solutions that deviate from core requirements. This combined assessment process demonstrably reduced an initial idea pool of over 70 candidates to a limited set of globally novel solutions, indicating an effective filtering mechanism for prioritizing innovative and relevant concepts.

Semantic clustering reveals that while a single-spurt approach to idea generation yields clusters of similar concepts, progressive provocation effectively diversifies ideas into distinct micro-clusters across both bird feeding and umbrella storage problem statements.

From Blueprint to Visualization: Bringing Concepts to Light

The ‘Scout’ and ‘Gatekeeper’ modules perform initial filtering of proposed concepts. ‘Scout’ evaluates ideas based on technical and practical feasibility, establishing constraints related to resource availability, current technological limitations, and potential implementation challenges. Simultaneously, ‘Gatekeeper’ assesses local novelty and diversity, preventing the reiteration of previously explored concepts within the defined scope and ensuring a breadth of innovative approaches are considered. This dual assessment process provides a focused set of viable and unique ideas for subsequent refinement and development, prioritizing concepts with both realistic potential and distinct characteristics.

The ‘Navigator’ module functions as a central synthesis engine within the conceptualization process. It receives input from both ‘Scout’, which provides feasibility assessments of proposed ideas, and the component deconstruction process, identifying reusable elements and potential building blocks. Utilizing this combined data, the ‘Navigator’ generates refined concepts, prioritizing those exhibiting both practical viability and novelty. This synthesis isn’t a one-time event; the ‘Navigator’ actively guides an iterative process, continuously adjusting and recombining components based on feedback and evolving constraints, thereby optimizing concepts for further development.

The ‘Director’ module functions as the conceptual architect, transforming validated ideas into formally defined concepts. This process utilizes the ‘PFIC’ (Properties, Functions, Interactions, Constraints) framework to meticulously detail features and implementation requirements. ‘PFIC’ facilitates a granular breakdown, ensuring all aspects of the concept are explicitly defined for development. Complementing this structured definition, ‘Leo’ generates photorealistic visualizations based on the ‘PFIC’ specifications, allowing for a clear and immediate understanding of the concept’s intended appearance and functionality. These visuals are integral for communication and further refinement throughout the iterative design process.

The Leo agent’s concept renderings for problem statements PS1-PS3 demonstrate a progression from structured PFIC concepts to photorealistic visualizations intended to facilitate final designer evaluation.

Beyond Ideation: A Holistic Evaluation Framework

The MIDAS system employs a comprehensive evaluation framework, termed ‘NDFR’, to move beyond simply generating ideas and instead prioritize truly impactful innovations. This framework assesses each concept along four key dimensions: Novelty, measuring originality and departure from existing knowledge; Diversity, ensuring a broad exploration of solution space and preventing incremental improvements; Feasibility, grounding concepts in practical constraints and resource availability; and Relevance, confirming alignment with defined goals and user needs. By systematically balancing these four factors, NDFR avoids the pitfalls of focusing solely on breakthrough ideas that may be impractical, or conversely, on safe but uninspired solutions; it fosters a portfolio of concepts that are both creative and realistically implementable, ultimately driving more effective innovation cycles.

MIDAS distinguishes itself through a uniquely collaborative design process, facilitated by the ‘Muse’ agent which integrates human insight directly into the innovation pipeline. This isn’t simply about humans vetting AI-generated ideas; rather, Muse actively participates in shaping and refining concepts alongside the artificial intelligence. By leveraging the strengths of both – human intuition, creativity, and contextual understanding coupled with the AI’s capacity for rapid iteration and exploration of vast datasets – MIDAS moves beyond automated generation. The system cultivates a participatory environment where human feedback isn’t a post-hoc evaluation, but an integral component driving the evolution of genuinely novel and impactful solutions. This synergistic approach allows for a more nuanced understanding of feasibility and relevance, ultimately fostering a cycle of continuous refinement unattainable through purely algorithmic means.

The MIDAS framework transcends basic ideation by establishing a dynamic loop of refinement and assessment, ultimately fostering more substantial creative outcomes. This isn’t merely about producing a high volume of concepts; it’s about continuously improving them based on rigorous evaluation criteria. Notably, the system achieved a level of semantic diversity so expansive that its generated idea pool was categorized as ‘noise’ by the DBSCAN algorithm – a clustering method that identifies outliers. This seemingly paradoxical result demonstrates the framework’s capacity to explore a truly vast and unconventional solution space, moving beyond incremental improvements to potentially groundbreaking innovations by deliberately resisting premature convergence on a limited set of options.

The MIDAS framework utilizes input prompts and multi-view embeddings to assign tasks to specific Large Language Models (LLMs), generating outputs that are then evaluated using the NDFR framework.

The pursuit of novelty, as demonstrated by MIDAS’s progressive ideation framework, echoes a fundamental tenet of system understanding. One must push boundaries to truly grasp limitations-and potential. Ken Thompson famously stated, “Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.” This sentiment applies directly to the core idea of the article; MIDAS doesn’t simply generate ideas, it iteratively refines them through a distributed agent network, effectively ‘debugging’ initial concepts to arrive at demonstrably more diverse and novel outcomes. The system’s success lies in its willingness to challenge, and ultimately improve upon, its own creations, embodying a relentless pursuit of refinement through rigorous testing-a process mirroring the very essence of discovery.

What’s Next?

The architecture presented here – a distributed intelligence dedicated to iterative refinement – feels less like a solution and more like a purposeful introduction of new questions. MIDAS demonstrates the capacity to generate diversity, but does that equate to meaningful novelty? The system excels at combinatorial exploration, yet the truly disruptive idea often originates not from rearranging existing elements, but from violating fundamental assumptions. One wonders if the ‘novelty’ observed is simply a measure of distance from the training data, a sophisticated echo rather than original thought.

The focus now shifts to the nature of the ‘progressive’ step. Is the system merely optimizing toward a locally defined peak, or is it genuinely exploring the broader conceptual landscape? Perhaps the most interesting limitation is not the AI’s capacity for ideation, but the inherent difficulty in evaluating its output. Human assessment remains the bottleneck – a subjective filter imposed on an ostensibly objective process. Future work must grapple with automating this evaluation, not by mimicking human judgment, but by defining entirely new metrics for conceptual value.

One begins to suspect that the real challenge isn’t building an AI that can think like a human, but one that can think differently. Perhaps the ‘bugs’ in MIDAS – the unexpected outputs, the illogical leaps – are not flaws to be corrected, but signals of a genuinely alien intelligence, a glimpse beyond the confines of our own cognitive biases. It may be that true creativity lies not in flawless execution, but in the beautiful, chaotic mess of unpredictable exploration.

Original article: https://arxiv.org/pdf/2601.00475.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/