MoFlow: Multi-Objective Agentic Workflow Generation Without Retraining

MoFlow is a method for generating agentic AI workflows that jointly optimizes multiple objectives such as accuracy, cost, latency, robustness, and consistency, instead of optimizing accuracy alone.

Existing methods optimize a single objective or a weighted sum, so each trained generator commits to one fixed trade-off and must be retrained when preferences change. MoFlow covers the Pareto front in one search and returns a workflow for any preference by lookup.

The paper is research-stage and names no open-source release or product, so direct adoption is premature. The practical value today is the idea of separating generation from preference selection, which is useful when designing agent systems that must balance cost and latency in enterprise settings.

It was evaluated against six strong baselines on six benchmarks spanning mathematics, code, and question answering, with baselines rerun for each testing preference.