why do some diverse teams outperform uniform ones while others fall apart, and does it depend on how the team talks to itself? We build agent-based models to find out.
Principal investigator: Tamás Kiss <kiss.t@wigner.hu>
Scope, questions, and goals
Management teams differ enormously in how their members’ skills are distributed: some teams are built from narrow specialists, others from generalists who can each do a bit of everything, and real teams sit all along that spectrum. A long line of organizational-behavior research asks whether this functional diversity helps or hurts a team’s ability to solve problems — and the honest answer, from both theory and experiment, is “it depends,” in ways that are hard to pin down from field data alone, since real organizations can’t be randomly assigned different skill mixes or communication rules to compare.
This project studies that question with agent-based simulation: teams of simulated agents, each with its own skill profile, are given tasks (also represented as skill requirements) to solve together, under different rules for how agents communicate and hand off work. That lets us treat diversity, task structure, and communication as independent dials we can turn one at a time — something no field study can do — and ask directly how they interact. Two established measures anchor the diversity side of this: DFD (dominant functional diversity — how much agents’ strongest skills differ from each other) and IFD (intrapersonal functional diversity — how generalist vs. specialist each agent is), both from Bunderson & Sutcliffe (2002).
What already exists
Our first results, published as arXiv:2604.21328 (“Role of diversity in team performance: the case of missing expertise, an agent-based simulation”), show that DFD and IFD don’t have a fixed effect on performance — depending on the team’s communication scheme and functional composition, more diversity can help or hurt, and the paper argues that a third measure, capturing a team’s aggregate expertise coverage, is needed alongside DFD and IFD to make sense of this. The simulation code behind this paper is openly released on CoMSES Net (model release, v1.0.0) — a ready starting point for a student project rather than a from-scratch build.
Planned future work
The next question is about communication structure itself. In practice, very different kinds of organizations pair specific task structures with specific communication patterns almost by convention — a military hierarchy passes information up and down a chain of command, while an academic research community communicates in a comparatively flat, all-to-all way. Is that pairing actually well-matched to the task, or just a historical habit?
This isn’t a new question — it goes back to a classic and still-surprising result in small-group research: Leavitt (1951) found that centralized, hierarchical communication networks solved a simple information-pooling task fastest, but Shaw (1954) found the opposite ranking on a task that required further reasoning over the pooled information, not just collecting it. That “which network structure suits which task” question has never fully settled, and it’s very much alive again today in the design of multi-agent AI systems: a 2025 survey of LLM-based multi-agent systems (Yan et al., 2025) notes that traditional communication architectures stop being sufficient as task complexity grows, and that current benchmarks still don’t measure the coordination properties needed to test which architecture actually fits which task.
We plan to extend our simulation framework — which already represents agents and tasks as skill vectors and gates who can talk to whom by skill similarity — to directly compare fixed network topologies (hierarchy, all-to-all, ring, star, small-world) and communication modes (agents passing tasks directly to each other vs. broadcasting work to a shared pool) against different kinds of task structure, and ask whether a team’s optimal communication design depends on its diversity profile in ways existing theory hasn’t tested with genuine, multidimensional skill differences between agents.
This is an active, hands-on line of work combining agent-based modeling, network science, and organizational behavior, well suited to a student project in MATLAB or Python — from extending the existing simulation to designing and running new experiments. If this sounds interesting, get in touch to discuss a topic.