Survival of the Useful: Evolutionary Boids as a Sandbox for Agent Societies

Xisen Wang, Qi Zhang

Keywords: Multi-Agent, Evolutionary Algorithm, Sandbox Simulation, boids

TL;DR: An sandbox simulation to test emergent intelligence through Boids algorithm

Abstract:

The emergence of complex intelligence from simple interactions has long fascinated artificial life and multi-agent research. Foundational work such as Boids showed how three local rules—cohesion, separation, and alignment—are sufficient to generate lifelike flocking without centralized control. In parallel, evolutionary algorithms explored how adaptation arises through variation and selection. Yet existing approaches remain limited: swarm models typically lack long-term adaptation, while evolutionary systems often converge prematurely and fail to capture emergent tool ecosystems.

We introduce TF-Boids: Survival of the Useful, a framework that unifies Boids-style local coordination with evolutionary selection in survival-driven environments. Each agent follows an observe–reflect–build loop to generate and refine tools, supported by automated testing, shared registries, and a Tool Complexity Index (TCI) that quantifies code, interface, and compositional sophistication. Local rules promote modularity and functional specialization, while evolutionary pressure retains strategies that enhance ecosystem robustness.

Our experiments span creative writing, data science, and research assistance domains, comparing Boids-enabled and baseline societies, and further incorporating evolutionary dynamics. Results show that Boids rules consistently reduce redundancy and favor compact, composable tools, while baseline systems trend toward heavier but more integrated pipelines. Evolutionary selection expands the ecosystem across generations, producing specialized tools with increasing capability.

This sandbox provides a tractable yet expressive platform for probing emergent intelligence through tool creation and refinement, with implications for multi-agent alignment, modular versus integrated design trade-offs, and the study of evolving ecosystems of intelligent agents.

Archival Option: The authors of this submission want it to appear in the archival proceedings.

Submission Number: 61