Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures

Summary

The largest structure-varying experiment currently available: 25,000 tasks across 8 models, 4-256 agents and 8 coordination protocols spanning externally imposed hierarchy through emergent self-organization, with a 44% quality spread between protocols (Cohen's d=1.86). Two results bear directly on the paper's design: a minimally scaffolded hybrid protocol beats centralized coordination by 14%, and the benefit of imposed structure is conditional on model capability - strong models self-organize while models below a capability threshold still need rigid structure. That interaction is the main threat to external validity for any fixed-structure claim and should be addressed explicitly, but note the study is single-authored and reports a composite quality score rather than separable epistemic dimensions.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a high-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule; organization theory applied to AI agents). Verified against: Opened https://arxiv.org/abs/2603.28990 (arXiv abstract page; confirmed title, single author, cs.AI, v1 30 Mar 2026, full abstract read)

Critique / open questions

Queued — not yet assessed.

Key quotes

How much autonomy can multi-agent LLM systems sustain -- and what enables it? (abstract)

We present a 25,000-task computational experiment spanning 8 models, 4--256 agents, and 8 coordination protocols ranging from externally imposed hierarchy to emergent self-organization. (abstract)