Pressure-Field Coordination for Multi-Agent LLM Systems

This repository contains the code and experiments for the paper:

Emergent Coordination in Multi-Agent Systems via Pressure Fields and Temporal Decay Roland R. Rodriguez, Jr. January 2026

Key Finding

Pressure-field coordination substantially outperforms all baselines on meeting room scheduling: 48.5% solve rate vs conversation-based dialogue (12.6%), hierarchical control (1.5%), and sequential/random baselines (<1%). Temporal decay is essential—disabling it reduces solve rate by 10 percentage points.

What's Here

crates/
├── survival-kernel/      # Core pressure-field coordination framework
└── schedule-experiment/  # Meeting room scheduling experiments and baselines
paper/
└── rodriguez-pressure-field-coordination-2026.tex  # Paper source (LaTeX)
results/
├── *.json                # Experiment result data
├── generate_figures.R    # Figure generation scripts
└── analyze_*.R           # Analysis scripts

Quick Start

# Build
cargo build

# Run tests
cargo nextest run

# Run experiments (requires Ollama with Qwen2.5 models)
OLLAMA_HOST=http://localhost:11434 cargo run -p schedule-experiment -- --help

# Quick test run (3 ticks, reduced trials)
QUICK=1 OLLAMA_HOST=http://localhost:11434 cargo run -p schedule-experiment

# Compile paper
cd paper && make final

The Idea

Traditional multi-agent LLM frameworks use explicit orchestration: planners, managers, message-passing. We take a different approach inspired by stigmergy (ant colonies, immune systems):

1. Shared artifact — agents modify a common workspace
2. Local pressure — quality signals guide greedy action
3. Temporal decay — prevents premature convergence
4. No communication — coordination emerges from shared state

Experiments

We evaluate on meeting room scheduling (270 trials across easy/medium/hard problems):

Strategy	Solve Rate
Pressure-field	48.5%
Conversation (AutoGen-style)	12.6%
Hierarchical	1.5%
Sequential	0.4%
Random	0.4%

Pressure-field achieves nearly 4× the solve rate of the next-best baseline. Effect size is large (Cohen's h = 1.07 vs conversation).

Requirements

• Rust 1.75+ (edition 2024)
• Ollama with Qwen2.5 models (0.5b, 1.5b, 3b)
• LaTeX distribution (for paper compilation)

License

MIT OR Apache-2.0