Multi‑agent LLM systems typically rely on generative models for expert execution, workflow planning, and artifact review. Yet many coordination steps only need bounded decisions, such as picking the next expert or judging artifact quality. Profiling three representative MAS frameworks reveals that routing and review consume 31.7%–43.5% of generative‑LLM tokens.
Replacing these steps with low‑cost decision models creates two interface gaps: a planning gap (generative planners output execution plans and hand‑off instructions only) and a review gap (reviewers provide actionable feedback that bounded decisions alone cannot supply).
Jev-LiteMAS bridges both gaps by coupling a lightweight decision model (Jev) with generative execution and feedback. To close the planning gap, Jev dynamically selects experts from the task description and prior expert outputs, using accumulated context to ground subsequent execution without a pre‑generated trajectory. To close the review gap, Jev screens intermediate artifacts against task criteria: passing artifacts proceed directly, while flagged ones are sent to an LLM for re‑review to decide acceptance or to generate corrective feedback.
Across four benchmarks and two backbone models, Jev‑LiteMAS reduces generative‑LLM token usage by 12.7%–54.6% compared to LAS, with accuracy loss within 1.34 percentage points, and lowers latency in seven of eight settings. Grounding routing in substantive expert outputs prevents up to 18.0 points of accuracy degradation.
Review: Jev‑LiteMAS achieves substantial compute and latency savings while preserving competitive task accuracy, demonstrating a viable path for decoupling decision‑making from generation in multi‑agent workflows.