New Method Automatically Designs LLM Workflows Without Manual Prompting

Building effective AI systems often requires stringing multiple large language models together in carefully orchestrated pipelines. Today's approach, however, depends heavily on human expertise to design these workflows and craft the prompts that guide each step—a labor-intensive bottleneck that slows deployment. A new research framework called FlowBot aims to remove this constraint by automatically learning optimal LLM workflows from data.

The Optimization Framework#

The core innovation treats workflow creation as a two-layer optimization challenge. The outer layer decides on the overall structure of the pipeline—which steps come first, how information flows between them, and where agent decisions should happen. The inner layer then fine-tunes the specific instructions for each individual LLM call, working through them one at a time. Rather than computing numerical gradients (as in traditional machine learning), both layers use textual gradients—feedback signals expressed as language that guide improvements to prompts and architectural choices.

How Textual Gradients Work#

The system propagates these language-based feedback signals backward through the pipeline, layer by layer. Each LLM component receives guidance on how to adjust its behavior to better serve the downstream tasks, much like how numerical gradients flow through a neural network. This modular approach allows the optimization to improve individual steps without rebuilding the entire system from scratch.

Performance Against Baselines#

According to the research, FlowBot-discovered workflows compete effectively against hand-designed pipelines and those generated through other automated methods. The system learns to build workflows that are comparable in quality to those created by human experts, but without requiring manual engineering for each new problem.

This work represents a step toward more automated AI system design, where machine learning can handle the labor of orchestrating complex multi-agent interactions.