A paper posted on arXiv addresses a trade-off in agentic workflows that use multiple large language model (LLM) providers. According to the abstract, assigning routine tasks to cheaper providers can lower generation costs, but switching providers may require the model to process accumulated context again.

The abstract goes on to suggest that keeping dependent tasks with the same provider can avoid this reprocessing. This points to a routing strategy that weighs cost savings against the expense of re-encoding context.

The source is truncated, so the full method and results are not available from this abstract alone. The paper is titled "Truthful and Expected-Payment-Optimal Model Routing in Graph-Based Agentic Workflows" and appears on arXiv.