Multi18

| Method | Avg. Reward (norm.) | Constraint Violations (%) | Cross-domain Transfer Gain | |----------|---------------------|----------------------------|----------------------------| | Mono | 0.61 | 22.1% | — | | Multi5 | 0.73 | 15.4% | +0.07 | | HRL | 0.69 | 18.9% | +0.04 | | Multi18 | | 8.3% | +0.21 |

Why 18? Empirically, we found that increasing the number of agents beyond 18 (e.g., to 24 or 32) led to diminishing returns and higher communication overhead ((O(n^2)) in graph edges). Below 12, the system underfit the diversity of constraints. The number 18 thus represents a “sweet spot” for mid-scale multi-domain problems—large enough to capture real-world heterogeneity, small enough for tractable coordination.

Multi18’s advantage was most pronounced in domains 14–18 (high regulatory strictness), where the arbiter prevented 94% of violations without aggressive reward shaping. multi18

Removing the coordination graph (i.e., independent agents) increased constraint violations to 27.4%, confirming the need for resource-aware arbitration. Reducing the context embedding to 8 dimensions hurt performance in the 10 text-based tasks (drop to 0.71 normalized reward), suggesting that 18 is a meaningful granularity for the tested diversity.

The “multi” prefix in AI often implies flexibility, but most multi-agent systems are tuned for 2–5 specific domains. We ask: Can a single architecture gracefully handle 18 qualitatively different environments without retraining? The number 18 arises naturally in certain industrial settings: 18 major languages, 18 time zones, 18 sub-components of a complex supply chain. We introduce Multi18—a proof-of-concept system where 18 specialized agents share a common communication protocol and a dynamic resource allocation mechanism. | Method | Avg

Real-world AI systems increasingly operate across multiple domains (e.g., healthcare, finance, logistics) while adhering to diverse constraints (e.g., legal, ethical, latency). We propose Multi18 , a modular framework designed for environments characterized by exactly 18 distinct operational modalities. The framework combines a lightweight negotiation protocol among specialized agents, a shared latent space for cross-domain state representation, and a constraint-satisfaction layer. Initial experiments in 18 simulated environments (varying resource availability and regulatory strictness) show that Multi18 reduces task-switching overhead by 37% and improves constraint adherence by 28% compared to monolithic baselines.

We introduced Multi18, a framework for multi-agent coordination across 18 distinct domains. By combining per-domain specialization with a global constraint-satisfaction layer, Multi18 outperforms monolithic and lower-agent-count baselines. The design principle of choosing N based on empirical complexity bounds (here, N=18) may generalize to other “multi-N” systems in applied AI. Below 12, the system underfit the diversity of constraints

Limitations: Multi18 assumes known domain boundaries and a static set of 18 environments. Extensions to open-ended domains (e.g., new domain appears online) remain future work.