Hack-Agents: A Multi-Agent System for Innovation - Proof of Concept and Implications for AI-Augmented Hackathons

Large Language Models (LLMs) are increasingly embedded in innovation processes, but uniform assistance can misalign with innovators’ shifting cognitive states, encouraging over-reliance and premature convergence that erode idea diversity. We introduce a phase-specific, multi-agent human-AI collaboration system that differentiates AI assistance across ideation, implementation, and evaluation. The system scaffolds phase transitions with: ideation feedback that probes assumptions and surfaces alternative framings to promote divergent exploration; an implementation chatbot that resolves technical uncertainties and automates routine steps to support convergent execution; and evaluation prompts that foster reflective iteration. We deployed the system in two innovation-focused hackathons: a pilot to improve robustness and user experience, and a follow-up study examining how phase-specific scaffolding shapes AI use and outcomes. Results show that phase-aligned assistance preserves idea diversity during ideation while still enabling efficient convergence toward feasible prototypes during implementation and reflection during evaluation. We derive design implications for AI-assisted innovation support tools.