From Human Crowds to Reliable Agent Swarms

Abstract: How can crowdsourced software development help us to design reliable swarms of cooperating software engineering agents? To answer this question first we need to understand the similarities between these two systems, which allows us to leverage empirical insights from human-driven systems to design more effective and resilient AI-driven ones. So, I begin with the empirical research I did on crowdsourced platforms, where large, heterogeneous populations participate in open-call marketplaces, solving tasks through a mix of competition and collaboration. Then I observe that AI swarm systems follow a similar pattern: autonomous agents operating in shared environments, decomposing problems and selecting actions based on local decision rules. Interestingly, despite the difference in agents, human versus artificial, both systems rely on decentralized coordination, where task allocation emerges from interaction rather than central control. Explaining performance in such systems is challenging, as outcomes depend on non-linear factors such as diversity, experience distribution, and task characteristics. However, by studying patterns in crowdsourced environments, such as participation dynamics, expertise effects, and failure behaviors, one can derive insights that directly inform the design and optimization of large-scale AI swarm architectures.

Short Bio: Raz Saremi is an industrial assistant professor at NYU and a staff research scientist at SERC (Systems Engineering Research Center). She has a Ph.D. in Systems Engineering from Stevens Institute of Technology, NJ, USA. Her research focuses on software engineering and decision sciences, in particular, Empirical Software Engineering.

Additionally, she has multiple years of experience in the machine learning and data science domain, primarily applied to the food and drug industry, where she has worked for companies such as ADM and J&J. Her work relies on a variety of tools, including neural networks, evolutionary algorithms, time series forecasting, network science, game theory, and Large Language Models.

Online link (Teams)