CrowdSeed AI

Inspiration In 2016, a landmark research paper proved that human crowds design significantly better robots when they can observe and iterate on each other's work ("Social" condition) compared to working in isolation ("Independent" condition). The social group unlocked a hidden "design quality" factor—largely driven by symmetry—that helped their robots walk farther. With the rise of agentic workflows, we asked a fundamental question: Does the "wisdom of the crowd" still hold true when the crowd is entirely made of AI agents? What It Does CrowdSeed AI is a simulation engine and interactive Streamlit platform that reruns this exact behavioral experiment using LLMs instead of humans. Agents are placed in a 5x5 grid environment and tasked with drawing connected lines to form a rigid robot body. The robots are then evaluated in a physics simulator to see how far they can walk. We split our AI agents into two distinct experimental tracks: - Independent Group (Control): Agents can only see and iterate on their own past design attempts. - Social Group (Experimental): Agents can view, adapt, and build upon the successful designs of their peer agents. How We Built It - Agent Infrastructure: Multi-agent LLM framework orchestrated to handle memory, and sequential design generations. - Physics & Evaluation Engine: A customized 2D physics evaluation layer that scores the distance traveled by each generated 5x5 grid robot. - Frontend & Analytics Dashboard: A Streamlit web application featuring results. What We Discovered (Our Real Results) Design Quality & Latent Factors: The Social AI agents scored significantly higher on the computed "design quality" factor than the independent agents (p = 0.0002, indicating a statistically significant result). Value & AI Infra Impact CrowdSeed AI provides the infrastructure to benchmark and study "artificial culture" and social synergy, proving that social connectivity makes AI agent swarms fundamentally more creative and capable.