Dynamic Latent Cloning with Continuous Barrier Guidance: Multi-Agent Vision-Predictor-Action

Vision-Language-Action (VLA) models emit actions token by token through decoders with billions of parameters. At batch size one, every decoding pass must bring all decoder weights on chip, so the latency of an action chunk is bounded below by N(|WLLM|b-Schip)/BW, and the closed-loop rates reported for such models are 1--5\,Hz. World models that reconstruct video face a second bottleneck: we show that any representation that reduces Gaussian pixel-nuisance distortion must carry nuisance information bounded below by a rate--distortion function, capacity that does not serve control. Multi-robot teams compound both problems. World models trained under centralized training with decentralized execution (CTDE) discard their joint model at execution, so agents that sample independently cannot break coordination symmetries. We present Continuous Multi-Agent Vision-Predictor-Action (CMA-VPA). A Plan--Realize JEPA master, running at 1--5\,Hz, forecasts macro-transitions over the set of agents and samples a joint mode; through dynamic latent cloning it retargets weight-identical clones with messages of at most 36 bytes, from which every clone reconstructs its conditioning deterministically using shared codebooks. Each clone runs a non-reconstructive JEPA predictor and a few-step conditional flow-matching solver whose sequential depth is $K+3$ network evaluations per decision, independent of the action horizon and of the team size. Safety enters the flow itself. CBF-DeGG-Flow specializes decoupled generative guidance for multi-agent flow matching (DeGG-Flow) to control barrier functions: each Euler step is guided by projecting the posterior-mean endpoint of the flow onto a polyhedron derived from a braking-consistent high-order control barrier function of relative degree two, and an executor projects every 10\,ms knot onto reciprocal per-agent constraints that carry explicit sampled-data, disturbance, and estimation margins, which extends the guarantees from the generative process to the executed commands. We prove forward invariance of the collision-free set under stated assumptions on disturbances, estimation errors, cooperation, and feasibility, with explicit local margins and a communication radius; a Lipschitz bound on the change between consecutive chunks under noise persistence and projection; soundness and full-dimensionality of the decentralized admissible input set, and a bound on how much the executed input deviates from the guided chunk; that symmetric policies and independently sampled passing sides stall in symmetric encounters, which a sampled joint mode resolves; a rate--distortion floor for the 36-byte bus; and a sequential-depth and roofline bound for an 18.7M-parameter clone (74.7\,MB at FP32). We pre-register software- and hardware-in-the-loop protocols, antipodal crossings of 4--16 quadrotors and collaborative dual-arm assembly, with hypotheses on decision latency below 15\,ms on a 15\,W Jetson Orin Nano, zero collisions under wind, a makespan reduction of at least 40% relative to serialized crossing, and latent messages of at most 40 bytes per cycle. No experimental results are reported.