SwarmWorld: Stigmergic technological evolution in societies of language-model agents

Recent advances in collective intelligence, such as ant colony optimization, particle swarm algorithms, and cellular automata, have demonstrated the power of local rules and environment feedback in generating complex behaviors. These methods highlight how simple local interactions can lead to global organization, yet often rely on predefined roles or centralized control. The emergence of large language models (LLMs) has opened new avenues for autonomous reasoning and communication among agents, enabling more flexible and scalable systems. Prior work includes projects like Generative Agents, Project Sid, and scientific swarms, which explore social behaviors, cultural transmission, and scientific hypothesis testing. However, these systems often depend on explicit communication, role assignments, or predefined workflows, limiting their capacity for open-ended, self-organized evolution. The gap lies in integrating environment-mediated knowledge transfer with scalable, role-free multi-agent systems to achieve durable technological progress.

The main challenge is to develop a decentralized, role-free multi-agent system capable of autonomous technological evolution, leveraging environment feedback as a primary mechanism. Existing systems struggle with scalability, robustness, and true self-organization, often requiring explicit communication or predefined roles. How can agents, initially homogeneous and without central control, build, inherit, and improve technologies solely through environmental cues? Addressing this question is crucial for advancing autonomous scientific discovery, materials design, and adaptive ecosystems. The difficulty lies in designing mechanisms that enable persistent knowledge storage, reliable inheritance, and resilience against disturbances, all while maintaining scalability and minimal supervision.

This work introduces SwarmWorld, a novel environment where homogeneous LLM agents operate without roles or predefined recipes, relying solely on environment feedback (stigmergy) for knowledge transfer. Key innovations include: 1) the spatial, persistent environment that stores artifacts and feedback, 2) the separation of cognition (agent proposals) from evaluation (deterministic simulator), 3) the integration of multiple mechanisms—communication, program inheritance, and stigmergy—to facilitate technological accumulation, 4) the demonstration that environment feedback alone can support complex, resilient societies. These innovations address limitations of prior role-dependent systems, enabling scalable, self-organizing technological ecosystems.

• �� Homogeneous LLM agents are placed in a spatial environment with fixed schemas for actions and materials. • Agents observe local environment states, retrieve memory, and generate structured plans for actions such as move, test, build, and code modification. • Actions are validated transactionally, with the deterministic simulator resolving physical consequences, including artifact functionality and environmental changes. • Agents can create, modify, and inherit programs and artifacts, which are stored physically in the environment, enabling persistent knowledge. • Different experimental conditions disable mechanisms like communication or program inheritance to isolate their effects. • The system tracks agent behaviors, artifact lineage, and performance metrics such as discovery frontier, resilience, and validated inventions over long simulation runs. • Statistical analysis compares populations of 50-200 agents across mechanisms, assessing diversity, robustness, and technological inheritance.

The experiments involve multiple conditions—full culture, no communication, no explicit culture, and isolated search—across populations of 50, 100, and 200 agents. Each runs for 800-3200 ticks, with repeated disturbance schedules to test resilience. Metrics include discovery frontier (best artifact performance), portfolio diversity, resilience under unseen disturbances, and number of validated inventions. The setup emphasizes long-horizon evolution, with artifact performance evaluated after agent removal to measure true technological robustness. Ablation studies isolate effects of mechanisms like communication and inheritance, while statistical analysis assesses the significance of observed differences. Results demonstrate environment feedback’s role in fostering diverse, resilient technological ecosystems.

Shared environments consistently outperform isolated search in producing broader, more resilient technology portfolios, with a 20-30% increase in diversity and robustness metrics. Agent behaviors naturally differentiate into exploration, construction, and coordination roles, with collaborative construction and inheritance pathways emerging without explicit prompts. Technologies evolve through physical artifact construction, program forks, and inheritance, with most reuse originating from physical observation rather than direct communication. Explicit cultural mechanisms enhance collaboration but are not strictly necessary for resilience. The system demonstrates that environment feedback alone can sustain complex, durable technological ecologies, validating the core hypothesis.

This framework can revolutionize autonomous materials discovery, robotic collaboration, and scientific hypothesis testing, especially in environments where explicit communication is limited. It supports the development of self-sustaining, scalable ecosystems capable of continuous innovation. In industry, it could enable autonomous factories or laboratories that adapt and evolve materials or processes without human intervention. In scientific research, it offers a new paradigm for distributed hypothesis generation and testing, reducing reliance on centralized control and predefined workflows.

The reliance on idealized physical simulations limits immediate real-world applicability; real environments are more complex and uncertain. Scalability is constrained by computational resources, necessitating algorithmic optimization for larger systems. The evaluation focuses on simulated performance; translating findings to physical materials and real-world systems remains a challenge. Future work must address environmental complexity, resource constraints, and validation through physical experiments to realize practical applications.