GraphMind: Self-Evolving Workflow Automation from Operational Traces
GraphMind is an end-to-end system for automating complex operational workflows without human input. It constructs action-centric workflow graphs from human resolution traces, then uses a multi-agent traversal engine with LLM reasoning to execute tasks. The Adaptive Traversal Reinforcement (ATR) mechanism enables self-optimization by reinforcing successful paths and decaying stale ones. This closed-loop approach addresses the challenge of adapting workflows over time, reducing reliance on manual configuration. The system was detailed in a paper on arXiv (2605.17617) as of May 2025.
Key facts
- GraphMind is an end-to-end system for workflow automation.
- It operates in three phases: offline graph extraction, online multi-agent traversal, and adaptive reinforcement.
- The system extracts workflow graphs from human resolution traces.
- It uses LLM-driven reasoning combined with graph-guided retrieval.
- Adaptive Traversal Reinforcement (ATR) reinforces successful paths and decays stale elements.
- The closed-loop mechanism enables self-optimization of the workflow graph.
- The paper was published on arXiv with ID 2605.17617.
- The system aims to reduce human effort in enterprise workflow automation.
Entities
Institutions
- arXiv