The Rabbit’s Foot Experiment

Overview

This experiment introduces a unique, non-consumable artifact — the Rabbit’s Foot — into the AgentFarm resource environment. Unlike standard resources that are gathered and consumed, the Rabbit’s Foot is a singular object that can be held by exactly one agent at a time. The holder receives a persistent RNG advantage that biases stochastic outcomes in their favor (combat rolls, gather yields, reproduction success). Agents can trade the Rabbit’s Foot to others via a new TRADE action, steal it through combat, or lose it on death.

The central research question: Will agents learn to retain the Rabbit’s Foot, and under what conditions does hoarding, trading, or ignoring it emerge as a dominant strategy?

Table of Contents

1. Motivation
2. Core Mechanics
3. Agent Interactions with the Rabbit’s Foot
4. RNG Advantage Model
5. Implementation Plan
6. Configuration
7. Metrics and Analysis
8. Experiment Plan
9. Hypotheses
10. Risk and Mitigation

Motivation

Standard AgentFarm experiments study emergent behavior around fungible resources — food that is gathered, consumed, shared, and regenerated. These resources are interchangeable and their value is linear. The Rabbit’s Foot introduces a fundamentally different economic object:

• Singular: Only one exists in the world at any time.
• Non-consumable: It is never depleted; holding it is costless.
• Transferable: It can change hands through trade or combat.
• Asymmetric: Its value is entirely positional — it benefits only the current holder.

This creates a game-theoretic landscape where agents must weigh the private benefit of holding the artifact against the social cost of being a target, and where trading it away can be either altruistic or strategic (e.g., trading it for resources when starving).

The experiment connects to broader questions about property, accumulation, and strategic valuation of unique goods in multi-agent systems.

Core Mechanics

The Artifact

The Rabbit’s Foot exists as a distinct entity tracked by the environment. It has three possible states:

1. On the ground at a grid position (can be picked up by any adjacent agent).
2. Held by an agent (provides the RNG advantage).
3. In transit during a TRADE action (instantaneous transfer to a neighbor).

Picking Up the Rabbit’s Foot

When the Rabbit’s Foot is on the ground, any agent within pickup_range (default: gathering range from config) can pick it up as part of a GATHER action. If multiple agents attempt to gather from the same position in the same step, the agent with the highest current resource level wins the tie (a proxy for strength/priority). Once picked up, the Rabbit’s Foot is bound to the holder until traded, stolen, or the holder dies.

Dropping the Rabbit’s Foot

The Rabbit’s Foot returns to the ground when:

• The holder dies (starvation, combat death) — it drops at the death position.
• The holder executes a TRADE action targeting a neighbor — it transfers directly (never hits the ground in a successful trade).

Agent Interactions with the Rabbit’s Foot

New Action: TRADE

A new action type is added to the ActionType enum:

class ActionType(IntEnum):
    DEFEND = 0
    ATTACK = 1
    GATHER = 2
    SHARE = 3
    MOVE = 4
    REPRODUCE = 5
    PASS = 6
    COMMUNICATE = 7
    TRADE = 8  # new

TRADE semantics:

The TRADE action enables agents to voluntarily give up the Rabbit’s Foot. Whether agents learn to do this — and under what circumstances — is a primary experimental outcome.

Combat and Theft

When an agent successfully attacks and kills a Rabbit’s Foot holder, the attacker automatically acquires the Rabbit’s Foot. This makes holders targets and creates a risk/reward tension:

• Holding the foot improves your RNG → you win more fights → positive feedback loop.
• But being known to hold it makes you a priority target → negative feedback from increased aggression.

If the attacker does not kill the defender (the defender survives), the Rabbit’s Foot stays with the original holder. There is no “pickpocket” mechanic — theft requires lethal combat.

Observation

The observation space is extended so that agents can perceive:

• Whether they hold the Rabbit’s Foot (binary feature in self-observation).
• Whether any visible neighbor holds it (binary feature per neighbor in the spatial observation).
• The ground position of the Rabbit’s Foot if it is within perception range and on the ground.

This information allows learning agents to make informed decisions about pursuing, trading, or avoiding the artifact.

RNG Advantage Model

The holder of the Rabbit’s Foot receives a luck multiplier applied to stochastic outcomes. The advantage is parameterized by a single float luck_bonus (default: 0.15, representing a 15% relative improvement).

The advantage is deliberately not applied to sharing, so that cooperation is not inflated by holding the artifact. The bonus affects competitive and self-interested actions, making the artifact a tool for individual advantage.

Tuning the Advantage

The luck_bonus parameter is a primary experimental variable. Sweeping it across values (e.g., 0.0, 0.05, 0.10, 0.15, 0.25, 0.40) lets us study how the magnitude of the advantage affects holding behavior:

• At luck_bonus = 0.0, the Rabbit’s Foot is inert — a control condition.
• At low values (0.05–0.10), the advantage is subtle; agents may not learn to value it.
• At moderate values (0.15–0.25), the advantage is meaningful; strategic behavior should emerge.
• At high values (0.40+), the advantage is dominant; the holder becomes extremely powerful, potentially creating runaway dynamics.

Implementation Plan

Phase 1: Rabbit’s Foot Entity and Environment Integration

Files to modify or create:

RabbitsFoot class sketch:

class RabbitsFoot:
    def __init__(self, position: tuple[int, int]):
        self.position = position
        self.holder_id: Optional[str] = None

    @property
    def is_held(self) -> bool:
        return self.holder_id is not None

    def pickup(self, agent_id: str) -> None:
        self.holder_id = agent_id
        self.position = None

    def drop(self, position: tuple[int, int]) -> None:
        self.holder_id = None
        self.position = position

    def trade(self, new_holder_id: str) -> None:
        self.holder_id = new_holder_id

Phase 2: TRADE Action

Files to modify:

trade_action() logic:

1. Validate agent holds the Rabbit’s Foot.
2. Find nearest agent within trade_range.
3. Transfer the Rabbit’s Foot to the target.
4. Optionally transfer trade_resource_cost resources from target to holder.
5. Log the trade event.
6. Return action result dict.

Phase 3: Luck Multiplier Integration

Files to modify:

The luck multiplier is read from the environment’s RabbitsFoot state, not stored on the agent — this keeps the agent stateless with respect to the artifact and avoids sync issues.

Phase 4: Observation Space Extension

Files to modify:

Phase 5: Configuration and YAML

Files to modify:

Phase 6: Metrics, Logging, and Analysis

Files to create or modify:

Configuration

# Added to default.yaml under a new top-level key
rabbits_foot:
  enabled: true
  luck_bonus: 0.15          # relative bonus to RNG outcomes
  trade_range: 30.0          # max distance for TRADE action
  trade_resource_cost: 0.0   # resources transferred from target to holder on trade (0 = gift)
  pickup_range: null          # null = use gather range from resource config
  initial_position: null      # null = random; or [x, y] for fixed start
  drop_on_death: true         # whether the foot drops when holder dies

Agent behavior weights (added to agent_parameters):

agent_parameters:
  SystemAgent:
    trade_weight: 0.05       # low propensity to trade away the foot
  IndependentAgent:
    trade_weight: 0.02       # very low — hoarding tendency
  ControlAgent:
    trade_weight: 0.10       # higher willingness to trade

Metrics and Analysis

Primary Metrics

Secondary Metrics

Visualization Ideas

• Timeline heatmap: Horizontal bar per simulation step, colored by which agent (or agent type) holds the foot. Gaps represent ground time.
• Trade network graph: Directed graph of all trades with edge weights for frequency.
• Holding duration distribution: Histogram of consecutive holding streaks, faceted by agent type.
• Luck bonus sweep plot: Holding duration and trade frequency as a function of luck_bonus.

Experiment Plan

Experiment 1: Baseline Characterization

Goal: Establish baseline simulation dynamics without the Rabbit’s Foot.

Collect standard metrics (dominance, cooperation, resource levels) for comparison.

Experiment 2: Default Rabbit’s Foot Introduction

Goal: Observe emergent behavior with a moderate luck bonus.

Key questions:

• Does holding duration increase over time within a simulation (do agents learn to hold)?
• Which agent type holds the foot the longest?
• How does the foot affect dominance outcomes compared to baseline?

Experiment 3: Luck Bonus Sweep

Goal: Map the relationship between advantage magnitude and holding behavior.

Key questions:

• At what luck_bonus threshold do agents begin to exhibit intentional holding behavior?
• Does a high bonus create runaway dynamics where the holder becomes unkillable?
• Is there an optimal bonus where interesting trading dynamics emerge?

Experiment 4: Trade Cost Sweep

Goal: Investigate whether requiring payment for trades changes behavior.

Key questions:

• Does a nonzero trade cost reduce trade frequency?
• Do agents learn to trade the foot when starving (trading luck for survival resources)?
• Does a high trade cost make the foot effectively non-tradable?

Experiment 5: Agent Type Composition

Goal: Test how population ratios affect Rabbit’s Foot dynamics.

Key questions:

• Does the most aggressive agent type (Independent) dominate foot possession?
• Does the most cooperative type (System) trade the foot away too readily?
• Does agent type composition change whether holding or trading is the dominant strategy?

Experiment 6: Learning Agent Focus

Goal: Test whether RL-trained agents learn to value and retain the Rabbit’s Foot.

Key questions:

• Do learning agents develop a preference for holding the foot over time?
• Does the learned policy converge to always-hold, always-trade, or conditional behavior?
• How many steps does it take for holding behavior to emerge?

Hypotheses

1. Holding will emerge: At luck_bonus >= 0.10, agents (especially learning agents) will develop longer average holding durations over the course of a simulation, indicating they learn the foot is valuable.

2. Independent agents will hold longest: Due to their lower cooperation propensity and higher aggression, Independent agents will accumulate more holding time than System or Control agents.

3. High luck bonus creates instability: At luck_bonus >= 0.40, the holder becomes so powerful that only combat death breaks their hold, reducing dominance switching and creating stagnant simulations.

4. Trade cost creates desperation trading: When trade_resource_cost > 0, agents will primarily trade the foot when their resources are critically low, exchanging long-term advantage for immediate survival.

5. System agents over-trade: System agents’ high sharing propensity will extend to the foot, causing them to trade it away more frequently than other types even when it would be beneficial to hold.

6. The foot amplifies existing dominance patterns: The agent type that is already dominant in baseline simulations will benefit disproportionately from the foot, as the RNG advantage compounds their existing strengths.

7. Ground time decreases over simulation: As agents learn (either through RL or through selection pressure from reproduction), the foot will spend less time on the ground in later simulation steps.

Risk and Mitigation

Summary

The Rabbit’s Foot experiment extends AgentFarm with a unique tradable artifact that grants an RNG advantage, creating a natural laboratory for studying possession, valuation, and strategic trade in multi-agent systems. The implementation touches the action system, observation space, environment state, and analysis pipeline, but is designed to be feature-gated behind a config flag so existing experiments are unaffected.

The six-experiment plan systematically varies the key parameters (luck bonus, trade cost, population mix, agent intelligence) to build a complete picture of how agents interact with a scarce, non-consumable resource — and whether they learn that holding on to luck is the smartest move.