Waggle: Learning One Anonymous Local Law for Self-Organizing LLM Swarms

Waggle is a multi-agent framework that shifts the learning target from explicit organizational structures (like roles or hierarchies) to a shared, anonymous local law. This reusable policy enables LLM agents to coordinate by making joint decisions on task actions, semantic communication, and local commitment updates based on bounded local views.

Key features of the Waggle approach include: Decentralized Organization: Coordination emerges and reorganizes online through the repeated execution of the local law, eliminating the need for global topology definitions or explicit role assignments. Swarm-Consistent Distillation (SCD): A training method that ensures the policy remains consistent across equivalent anonymous views and learns to predict evolving local coordination states, improving stability and reorganization after counterevidence. Scalability and Transfer: The same learned law remains effective as population sizes and interaction budgets change, retaining over 96% of substrate-specific oracle quality and transferring across different coordination settings without retraining. Anonymous Interaction: Agents use opaque, episode-local handles rather than persistent identities, ensuring that coordination behavior is invariant to agent renaming or local reference permutations.

Waggle addresses coordination in LLM swarms by asking whether a single, anonymous local coordination law can be learned and reused across agents, rather than specifying roles, communication graphs, or task-specific orchestrators. The central idea is to treat coordination as a reusable local policy: each agent observes a compact representation of its neighborhood or shared context and applies the same learned rule to decide how to act, defer, aggregate, or update. Because the law is anonymous and local, agents remain interchangeable and the swarm can be deployed over changing topologies without re-architecting the system.

The key contribution is a shift from bespoke multi-agent designs to a shared coordination primitive. Instead of engineering role assignments, hand-written prompts, or fixed workflows for each application, the paper proposes learning a transferable law that encodes when and how agents should coordinate. This is potentially important because LLM swarms are often brittle under scale and reconfiguration: adding, removing, or reassigning agents can require redesign. A reusable local law offers a more modular path to self-organization, making it easier to deploy large populations of similar agents in dynamic environments.

It matters for systems research because it connects swarm intelligence—where global behavior emerges from simple local rules—with modern LLM agent systems. If such laws can be learned robustly, they could reduce coordination overhead, improve generalization across tasks and topologies, and enable more decentralized, fault-tolerant agent collectives. The broader insight is that effective swarm behavior may not require many specialized roles; it may be possible to instantiate a single anonymous coordination law and let distributed agents produce useful collective behavior.