FractalSwarm

Multi-agent systems can distribute work across many agents, but what happens when the coordinators managing those agents suddenly disappear? FractalSwarm explores a different approach to multi-agent reliability: a hierarchical swarm where coordination authority can be recovered deterministically, without relying on an LLM to elect a new leader. When a coordinator fails, surviving coordinators independently compute the next heir using rendezvous hashing and the recorded event log. Agents in unaffected branches can continue working, while successors recover ownership and adopt results that arrive from previously assigned work. This helps avoid unnecessary redispatch and keeps the system moving through coordinator failures. Built with Meta AI models, FractalSwarm combines Muse Wonder for planning, Muse Spark for task execution and information extraction, and Llama PromptGuard 2 for an additional security layer. Its architecture also includes event logging, deterministic replay, ownership checks, and failure-injection experiments. We evaluated it against durable supervisor baselines across repeated crash scenarios. The results show a specific strength: FractalSwarm retains more active sessions when coordinators fail, while matching an ideal durable supervisor on some recovery measures—not outperforming it across every metric. The project is built around one question: Can a multi-agent system recover its coordination structure without stopping the work already in progress?