Multi-agent LLM systems typically lean on generative models for everything, including expert execution, workflow planning, and artifact review. But a new arXiv preprint argues that many coordination operations—such as picking the next expert or judging an artifact—are actually bounded decisions, not open-ended generation tasks.
The proposed framework, Jev-LiteMAS, separates these decisions from generation. By treating coordination as a decision problem rather than a text-generation problem, the approach could let lighter-weight models handle orchestration while reserving generative LLMs for tasks that genuinely require them.
The abstract is truncated in the source, so implementation details and experimental results are not covered here. The core idea, however, is clear: not every step in a multi-agent workflow needs a full generative pass.