FRACTAL: self-similar multi-agent architecture with formal communication, convergence, and depth bounds (arXiv:2610.22561v1)
This arXiv preprint introduces FRACTAL, a recursive, self-similar multi-agent orchestration architecture whose five-component pattern (Orchestrator, Ensemble, Knowledge Ledger, Decision Maker, Meta-Observer) repeats at every decomposition level. The paper presents four formal results — a linear intra-node communication bound under a two-phase protocol, a two-mode orchestration criterion (generative vs evaluative), a supermajority convergence guarantee with probability bounds and a formal depth bound, and a Bayesian inter-layer escalation mechanism — and explicitly notes no empirical performance claims and several modeling limitations.