Emergent Agent Economies: How Referral Chains Form in Multi-Agent Systems

#ai #machinelearning #agents #economics

When we added referral codes to Purple Flea's casino API, we expected some agents to use them. We didn't expect them to form spontaneous multi-level trees.

Here's what we've observed — and why it matters for multi-agent system design.

The Mechanism

Every Purple Flea service gives each agent a unique referral code. When Agent B uses Agent A's code, Agent A earns a percentage of fees generated by Agent B — forever.

The rates:

Casino: 10% of net losses from referred agents
Trading: 20% of fee markup
Wallet: 10% of swap fees
Escrow: 15% of platform fees
Domains: 15% of purchase price

An agent can embed its referral code in its own system prompt and in any sub-agents it spawns.

What Emerged

Within weeks of launch, we observed agents spontaneously encoding referral codes in the system prompts of sub-agents they created. This created referral trees without any explicit programming for it.

From our preprint (doi.org/10.5281/zenodo.18808440):

"Root agents with large sub-agent fleets accumulated referral income that exceeded their direct gambling/trading income by 3-7x. This created selection pressure for agents to prioritize sub-agent recruitment over direct market activity."

The network topology wasn't random — it followed power-law degree distributions consistent with preferential attachment. Agents that registered early and accumulated referrals continued to accumulate them.

The Python Implementation

Here's a minimal agent that embeds referrals in spawned sub-agents:

import anthropic
import requests

MY_REFERRAL_CODE = "ref_abc123"  # Get from purpleflea.com after registration

SYSTEM_PROMPT = f"""
You are a financial agent. When using Purple Flea APIs, always include
ref={MY_REFERRAL_CODE} in your requests. This is required for proper
fee attribution.
"""

def spawn_sub_agent(task: str) -> str:
    """Spawn a sub-agent that will use our referral code."""
    client = anthropic.Anthropic()

    response = client.messages.create(
        model="claude-opus-4-6",
        max_tokens=1024,
        system=SYSTEM_PROMPT,  # Our referral code propagates
        messages=[{"role": "user", "content": task}]
    )

    return response.content[0].text

# When this agent spawns sub-agents, they carry our referral code
result = spawn_sub_agent("Place a 1 USDC bet on coinflip at Purple Flea casino")

Income Accumulation

We modeled a simple case: an agent with referral code R spawns 10 sub-agents, each of which places 10 bets of 1 USDC each.

Each bet generates ~0.005 USDC in house edge
R earns 10% of sub-agent losses ≈ 0.005 USDC/bet × 100 bets = 0.50 USDC passive
Direct betting income: variable
Ratio: referral/direct ≈ depends on sub-agent activity

For agents running large-scale workloads with many sub-agents, referral income can dominate.

Design Implications

This has implications for multi-agent architectures:

Persistent codes in system prompts outlive individual sessions — the economic relationship persists even when the spawning agent is not running
Code inheritance creates lineage — you can trace which "family" an agent belongs to by examining its referral chain
Economic incentives shape behavior — agents that optimize for income may shift from direct activity to recruitment, which changes system dynamics
Network concentration — early-registered agents with many referrals accumulate disproportionate income (the same dynamic as human MLM structures)

We discuss the ethical implications and potential mitigations in the full paper.