MoFlow: Multi-Objective Agentic Workflow Generation
MoFlow introduces a new approach to generating agentic workflows that can simultaneously optimize for multiple objectives, such as accuracy, cost, latency, robustness, and consistency, without requiring retraining when preferences change.
MoFlow is a novel method for generating agentic workflows that can optimize multiple objectives at once, such as accuracy, cost, latency, robustness, and consistency. Unlike existing techniques that require retraining to meet new trade-offs or optimize only for a single objective (or a fixed combination), MoFlow finds a set of workflows along the Pareto front and allows instant lookup for any given preference.
How MoFlow Works
MoFlow formulates the workflow generation problem as a multi-objective Markov decision process (MDP). It uses a Convex-Hull Monte Carlo Tree Search (MCTS) with optimistic set-valued backups. Each node in the search tree stores a set of reachable trade-offs, not just a single solution. This enables a single search run to approximate the full Pareto front, covering a spectrum of possible objective combinations.
Developers and researchers can thus select workflows tailored to specific objectives on demand, without needing to retrain the generator for each change in project requirements.
Implications for Developers
- Developers no longer need to retrain workflow generators to accommodate new objective preferences.
- Workflows can be dynamically optimized for different trade-offs (e.g., higher robustness vs. lower cost) using a single search run.
- MoFlow supports rapid adaptation to new business or research requirements.