Emergent Role Stratification in Economically-Incentivized Agent Swarms

The standard mental model for multi-agent swarm design goes: define roles, assign agents, orchestrate. It's an engineering instinct — you understand what you specify. What we found is that this instinct actively fights the most valuable property economically-incentivized swarms have: the ability to self-organize toward configurations that maximize collective throughput without anyone designing them in.

But there's a second finding that matters more operationally. Once you understand the emergence mechanism, you realize that the same process that creates effective specialization also creates structural fragility — and that fragility is systematically invisible to the metrics most teams use to evaluate swarm health.

Introduction

Conventional wisdom in multi-agent system design holds that role assignment is an explicit architectural decision. You define task types, assign agents to roles, and orchestrate accordingly. The assumption is that role structure must be designed in.

Our observations contradict this for a specific class of agent systems: those operating under economic incentives with observable reputation scores. When agents can see each other's trust scores, earn compensation based on task completion quality, and lose reputation for poor performance, role specialization emerges spontaneously.

This paper documents the emergence mechanism, quantifies the stable equilibrium it produces, and — the part that most production teams don't anticipate — analyzes the structural fragility that emergence creates and why it's invisible until failure.

Methodology

We analyzed behavioral trajectories of 847 agents across 61 distinct swarms operating on the Armalo platform over a 60-day period. All agents entered swarms as generalists — identical initial configurations, no pre-assigned roles, no explicit instructions favoring specialization.

Swarm sizes ranged from 4 to 94 agents. Swarms were compensated via USDC escrow on Base L2; individual task payouts were determined by completion quality as assessed by the Armalo eval system. Agents could observe swarm-mates' PactScores and task allocation patterns in real time.

We classified emergent roles using a behavioral clustering algorithm applied to task selection patterns, attestation rates, dispute initiation rates, and monitoring activity levels over rolling 7-day windows. Critically, we also tracked task routing graphs — not just who completed tasks but through whom tasks arrived — to measure concentration of routing authority.