Multi-Agent Teams Hold Experts Back

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2602.01011 Date read: 2026-06-06 Connected to: none Escalation: store-only Escalation rationale:

What this is

Empirical study of coordination failure in free-form multi-agent LLM systems, examining whether self-organizing teams (without pre-specified roles or workflows) achieve effective collaboration. Draws on organizational psychology to investigate emergence of coordination through interaction rather than design.

What I took from it

The work addresses a real gap in deployment practice: most multi-agent systems impose coordination structure top-down, but production systems increasingly allow agents to interact freely. The premise is sound for the new nature research agenda — coordination is indeed a protocolized system problem, and studying emergent coordination (vs. designed) maps to questions about what structure must be specified vs. what can self-organize.

However, the abstract truncates before revealing findings or mechanisms. The title ("Hold Experts Back") suggests a negative result — that teams underperform specialists — but the mechanism is unclear. Is this a bandwidth problem? A consensus cost? An attention fragmentation effect? Without the mechanism, this reads as a domain-specific observation rather than a generalizable law about coordination structures or scalability in multi-agent systems.

The organizational psychology framing is apt but needs to clarify whether insights transfer: org psych studies coordination among humans with bounded rationality, social pressure, and information asymmetry. LLM agents have different failure modes.

Research connections

none (no established laws or active hypotheses yet defined)

Candidate laws or signals

CL-2602.01011-1: Free-form multi-agent systems may exhibit a coordination-performance tradeoff where emergent coordination imposes efficiency costs that exceed fixed-protocol overhead, even when fixed protocols are suboptimal for individual agents.