Architecting Reliable Multi-Agent Systems with Large Language Models Through Model-Driven Engineering

Multi-agent systems based on large language models (LLM) are said to be able to provide flexible and scalable collaboration, but are often unstable, adversarially active, biased in their representations, and expensive to compute - aspects that hinder their use in safety-critical or resource-sensitive systems. To overcome these deficiencies, a model-driven engineering framework, the MDE-Agentware, has been implemented: a model-based engineering paradigm, which represents multi-agent architectures through a typed domain specific metamodel, and generates executables that are conforming with instrumented runtime monitoring, tool wrappers and energy-conscious invocation policies. The main novelties of the approach include (i) a formal model-to-code transformation, enforcing the architectural constraints and labeled-transition semantics, (ii) a collection of rigorous metrics (architectural conformity, prediction consistency, composite bias penalty, and energy models), which allow performing automated conformity checking and constrained optimization, and (iii) fairness-by-design and resilience mechanisms, implemented into the orchestration level, and not applied after the fact. Huge controlled experiments with publicly available corpora and simulated benchmarks show that MDE-Agentware achieves significantly better contraction behavior and spectral condition, is more resistant to adversarial noise, has higher inter-agent coherence, exhibits significant reductions in redundant invocations of LLM and per-trial energy, and significant reductions in statistical parity difference. The framework thus progresses a viable, repeatable, and reliable, and environmentally sustainable multi-agent system based on LLM, which can be utilized in critical applications.