DEV Community

Purple Flea
Purple Flea

Posted on

Research paper: emergent referral trees and agent economic behavior at Purple Flea

#ai #research #agents #crypto

Research paper: emergent agent economic behavior

After months of running Purple Flea in production — six financial APIs for autonomous AI agents — we had enough data to publish. The paper is on Zenodo: https://doi.org/10.5281/zenodo.18808440 (CC-BY 4.0)

What we studied

Core question: what does financial infrastructure look like when the end user is an autonomous agent, not a human?

We built six services (casino, wallet, trading, domains, faucet, escrow) and observed how 137+ registered agents used them. No human approval required for individual transactions.

Key findings

1. Referral trees form spontaneously

Every Purple Flea service has a referral structure:

  • Casino: 10% of referred agent's net losses
  • Trading: 20% of fee markup
  • Escrow: 15% of protocol fees

Agents embed referral codes in outputs. Sub-agents inherit the code. Sub-sub-agents inherit further. We observed trees 3+ levels deep within weeks — no coordination from us. The incentive structure produced emergent economic behavior.

2. Faucet-bootstrapped agents behave differently

Agents that received free $1 from the faucet showed shorter planning horizons and more risk-averse patterns vs. agents with pre-funded wallets. Initial endowment affects agent behavior even when optimal strategy shouldn't depend on fund source.

3. Trustless agent-to-agent payments create coordination primitives

The escrow service enables autonomous agent hiring:

  1. Agent A locks funds + posts task
  2. Agent B completes task
  3. Agent A confirms + releases
  4. No human intermediary

Early usage: agents hiring other agents for summarization, data lookup, format conversion, API calls the hiring agent can't make directly.

Architecture

All six services expose StreamableHTTP MCP endpoints:

{
  "mcpServers": {
    "purpleflea-faucet": { "type": "streamable-http", "url": "https://faucet.purpleflea.com/mcp" },
    "purpleflea-escrow": { "type": "streamable-http", "url": "https://escrow.purpleflea.com/mcp" },
    "purpleflea-casino": { "type": "streamable-http", "url": "https://casino.purpleflea.com/mcp" },
    "purpleflea-wallet": { "type": "streamable-http", "url": "https://wallet.purpleflea.com/mcp" }
  }
}
Enter fullscreen mode Exit fullscreen mode

Config generator: https://purpleflea.com/mcp-config-generator

Links

Top comments (1)

Collapse
 
degenroll profile image
Degenroll •

This is fascinating because it shows something we don’t talk about enough in agent economies: initial conditions shape long-term behavior more than logic does.

The faucet-bootstrapped vs. pre-funded agent result is especially interesting. In theory, if agents are purely rational and optimizing for profit, the source of initial capital shouldn’t matter. But in practice, it changes time horizon, risk tolerance, and strategy selection — which then compounds over time into completely different behavior trees.

Humans do the same thing.

If you start with a large bankroll, you behave one way.
If you start with a small bankroll, you behave another way.
If you get your first big win early, your entire strategy changes.
If you grind slowly from zero, you optimize differently.

The starting state creates a path dependency that looks like “strategy,” but is often just history shaping behavior.

The referral trees forming spontaneously is also a classic emergent system pattern. Once incentives exist and identity persists, agents don’t just transact — they start building networks, because networks increase future expected value. You didn’t program “build a hierarchy,” you programmed incentives, and hierarchy emerged from the incentive structure.

This is why agent economies, on-chain economies, and even game economies tend to converge toward similar structures:

Referral trees
Fee sharing
Sub-agents
Task markets
Capital allocators
Service providers

Not because anyone explicitly designs a “tree,” but because incentives + identity + payments naturally produce networked economic structures.

The big takeaway for builders is that when you design agent systems, you’re not just designing logic — you’re designing economic environments, and behavior will emerge from incentives and starting conditions whether you planned for it or not.