A Preliminary Exploration of Evolving Agent Societies through Simple Local Rules

Xisen Wang, Qi Zhang

Keywords: multi-agent systems; collective intelligence; emergent coordination; Boids model; tool-building agents; evolutionary algorithms; decentralized automation; local interaction rules; AI ecosystems

TL;DR: Simple local rules combined with evolutionary feedback can yield scalable, self-organizing agent societies that build and refine tools without predefined workflows.

Abstract: How much collective intelligence can emerge from simple, decentralized rules without heavy, predefined workflows? We explore this question through a framework that couples Boids-style local coordination with explicit evaluation and selection in a survival-driven, tool-building ecology. Agents interact via three local rules—cohesion, separation, and alignment—and follow an observe–reflect–build loop to generate and refine tools within an ecosystem that includes automated tests, shared registries, and a Tool Complexity Index capturing code, interface, and compositional sophistication. Positioned as a preliminary study, this work treats evolution as a complementary lens: local rules catalyze collaboration and modularity, while selective feedback favors strategies that persist across generations. Across text analysis, data science, and simulation modeling tasks, evolutionary-Boids societies increase throughput and balance contributions among agents while maintaining reliability, though current prompting tends to suppress deep tool composition. The resulting systems produce more and self-contained tools rather than extended pipelines, suggesting a breadth-first mode of exploration in its current setup. Overall, the framework offers an early step toward understanding how simple, local interactions and evolutionary pressure together shape the emergence of organized, evolving agent ecosystems.

Submission Number: 197