Work

Beyond Generative Agents: Can Heterogeneous Intelligence Produce Wiser and More Stable Multi-Agent Systems?

Year: 2026
Type: preprint
Abstract: Multi-agent artificial intelligence is frequently discussed as a matter of multiplying language-model roles. Yet an operational intelligence system can draw on quite different computational pathways: generative natural-language models, predictive machine-learning or deep-learning models, human-authored programs, and machine-generated programs. This conceptual paper examines whether a dynamic network composed from these four pathways can support decisions that are more reliable, stable, explainable, and appropriately governed than simpler arrangements. The four pathways are not mutually exclusive classes of agency; rather, they separate mechanisms and software provenance that may be realized as autonomous agents, bounded tools, or conventional services. Drawing on heterogeneous knowledge systems, neuro-symbolic AI, adaptive coalitions, evidence provenance, assurance cases, selective prediction, and software verification, the paper proposes a claim-qualified coalition method. Its central design question is whether an additional autonomous boundary preserves independently consequential evidence, state, authorization, or recovery that a resource-matched simpler system cannot preserve. A set of constructed adversarial cases and prospective comparison arms specifies how this question might be falsified. No new universal agent architecture, mathematical theorem, executed benchmark, empirical improvement, or measured form of wisdom is claimed. The contribution is a bounded comparative research framework for deciding when heterogeneous computational capabilities should be composed, separated, reduced, or withheld from action.
Language: English
Open Access status: green
OpenAlex ID: W7234735868
Cites:
Cited by:
Related to: 10