WorkflowGen: Trajectory-Driven Workflow Generation for LLM Agents
WorkflowGen has launched an innovative framework aimed at streamlining the process of creating workflows for large language model (LLM) agents. This system focuses on a flexible, experience-based trajectory approach. Unlike conventional methods that start from scratch for every request—resulting in excessive reasoning demands, high token consumption, inconsistent results, and no reuse of past experiences—WorkflowGen addresses these issues by capturing full execution paths early on and pulling valuable insights from them. These include error patterns, ideal tool associations, parameter structures, execution routes, and exception avoidance strategies. By employing a closed-loop system, it generates efficient outputs based on variable nodes using techniques like trajectory rewriting and experience updating. The aim is to boost efficiency and success rates in complex tasks, such as business inquiries and workflow management. You can check out the study on arXiv under the identifier 2604.19756.
Key facts
- WorkflowGen is an adaptive, trajectory experience-driven framework for automatic workflow generation.
- It addresses high reasoning overhead, excessive token consumption, unstable execution, and inability to reuse past experiences in LLM agents.
- Traditional methods generate workflows from scratch for every query.
- WorkflowGen captures full trajectories early in execution.
- It extracts reusable knowledge at both node and workflow levels.
- Extracted knowledge includes error fingerprints, optimal tool mappings, parameter schemas, execution paths, and exception-avoidance strategies.
- It employs a closed-loop mechanism with trajectory rewriting and experience updating.
- The framework reduces token usage and improves efficiency and success rate.
- The paper is on arXiv with identifier 2604.19756.
Entities
Institutions
- arXiv