Snake Swarm: A Heterogeneous Multi-Agent Architecture for Reliable LLM-Based Task Execution

Abstract

Recent advances in large language models (LLMs) have not eliminated their tendency to produce errors on complex reasoning tasks. Existing mitigation strategies---scaling to larger models, majority voting, or post-hoc repair---each carry fundamental limitations in cost, reliability, or information fidelity. We present Snake Swarm, a multi-agent architecture that reduces system-level error rates by orchestrating heterogeneous model redundancy with structured arbitration, without requiring improvements to any individual model. The architecture models tasks in a four-dimensional space (value, time, domain, difficulty) and routes them through configurable execution pipelines composed of a structured-plan head, a dual-output body, and a recording tail. A three-layer arbitration system enforces at least one heterogeneous judge, conducts deep-dive reasoning-chain review triggered by dissenting opinions, and subjects judges themselves to meta-level accountability through tracking, anti-consensus protection, and observation periods. In preliminary experiments on six test cases, the heterogeneous three-judge configuration achieved an 83% error detection rate, compared to 50% for three same-model judges and 17% for a single judge---suggesting that heterogeneous arbitration constitutes a qualitative threshold rather than an incremental improvement. The architecture's core hypothesis---that structured heterogeneous review can outperform a single larger model at equal token budget---remains under experimental verification.