Consensus algorithms for a multiagent operating system
Edge multiagent systems call for operating-system-level agreement primitives that tolerate peer failures, lossy and unsynchronized messaging, and heterogeneous local state, rather than relying on a permanent timing master. Neighbor exchange is modeled as a time-delay multi-agent system under bounded communication delays. This article presents two complementary building blocks—monotone logical-time synchronization via max/SoftMax consensus with PI-style implementation, and consensus-based accelerated distributed simultaneous perturbation stochastic approximation (ADSPSA) for cooperative tracking under zeroth-order oracles— together with a multimodal composition principle: logical-time and spatial coordinates are stacked per agent, coupled through multimodal losses and a block preconditioner on the ADSPSA channel. Formal convergence analysis for the logical-time synchronization primitive is developed in companion works on monotonic logical time (an accepted ApPLIED 2026 contribution and a manuscript submitted to IEEE access); fully coupled multimodal validation—joint logical-time and tracking experiments under delays and losses—remains outside the scope of this paper. For the Ψ-preconditioned recursion we develop a majorization framework that yields an asymptotic upper bound on the tracking-error covariance under explicit smoothness, centering, and spectral-contraction assumptions. Numerical experiments on a heterogeneous quadratic surrogate—isolating the tracking block but using a time/tracking block structure for Ψ—illustrate how an empirically tuned preconditioner can lower the asymptotic error level relative to the identity map, sometimes at the cost of slower transient decay.