LLM agentic workflows often assign different models to specialized roles, but finding good assignments normally requires evaluating many candidates. A new paper from researchers at MIT, Columbia, and other institutions introduces Cost-Coupled Gittins (CC-Gittins), a method that recommends Pareto-optimal assignments under quality, cost, and optionally latency trade-offs. The key insight is that evaluation and deployment costs are coupled: assignments that cost more to deploy also cost more to evaluate, so each evaluation can inform both objectives simultaneously.

The algorithm alternates between seeking higher quality and lower deployment cost, reusing every evaluation to update both searches. After each batch, it recommends assignments that are nondominated under uncertainty-adjusted estimates of quality and cost. In experiments across eight quality–cost matrices from workflows with two to five roles and up to 1,681 candidate assignments, CC-Gittins recovered a substantial fraction of the true Pareto frontier at roughly 10% of the exhaustive-evaluation cost, and most of the frontier with few false positives by 30%.

The authors also tested latency as a third objective, showing the method extends beyond two objectives. The paper is a preliminary version accepted at the NeurIPS 2026 Workshop on Resource-Aware Agentic AI, so the results have not yet been peer-reviewed in a full venue. Still, the approach offers a practical way to cut the search cost of configuring multi-model agentic systems without sacrificing the ability to explore quality–cost trade-offs.