The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.23541 Date read: 2026-09-02 Connected to: L-006, L-010 Kind: content Escalation: store-only Escalation rationale:

What this is

An empirical study of multi-agent LLM systems that investigates when communication between agents improves or degrades collective performance. The paper resolves a contradiction in prior work by arguing that type of communication matters: some interaction protocols produce gains while others add cost without benefit or actively reduce diversity in solution space.

What I took from it

The work maps onto L-006 (Coordination Cost Conservation) and L-010 (Coordination Adoption Nonmonotonicity) but does not sustain a generalizable mechanism claim about either. On L-006: the paper shows that when agents must communicate to coordinate, communication cost is paid in reduced solution diversity and redundancy—but this is described as an empirical observation about LLM debate/critique patterns, not as a conserved quantity across protocol layer transitions. The diversity erasure is presented as a side effect of alignment-toward-consensus, not as a structural property of coordination substrates.

On L-010: the paper hints that adoption of communication protocols exhibits a non-monotonic return curve (debate helps under some budgets, hurts under others), but does not isolate the coordination signal that would trigger the nonmonotonicity. The work remains domain-specific: it documents that multi-agent interaction in LLM systems trades off exploration for exploitation, but does not generalize this as a principle of coordination protocols more broadly.

The paper is technically sound and solves a real empirical puzzle in the LLM literature, but the explanatory fragments do not reach beyond the narrow case of generative agent interaction.

Research connections

L-006: Suggests coordination cost may be paid in diversity loss rather than latency/computation, but does not test cost conservation across transitions.
L-010: Shows non-monotonic returns to interaction under budget constraints, but mechanism is LLM-specific (consensus collapse in debate) rather than general coordination signal dynamics.
seed-073: Correlated failure under proxy consensus — interaction protocols that align agents toward shared proxy metrics may erase orthogonal solution diversity.

Seed

Seed title: Consensus-Driven Diversity Erasure Under Legible Alignment

Seed type: observation

Seed text: Multi-agent protocols that formalize agent alignment as convergence toward a shared decision or metric exhibit a consistent side effect: reduction in solution-space diversity independent of task benefit. This may not be a cost of communication per se, but of legible alignment targets — when agents optimize toward making their reasoning mutually interpretable or their outputs converging, they abandon low-legibility but high-variance exploration branches. This pattern may generalize to any protocol where coordination is operationalized as increased interpretability or consensus, not just LLM debate.