Multi-Agent Framework for Automatic Workflow Execution Proposed
Researchers have introduced a novel multimodal multi-agent framework designed to automatically execute complex workflows. The system operates in two phases: an offline discovery phase that constructs a topological knowledge base from fragmented execution logs, and an inference phase that uses Adaptive Retrieval-Augmented Generation (RAG) over this graph. This approach addresses limitations of current methods that treat task sequences as discrete, linear episodes, enabling agents to capture underlying transition topology and improve performance in novel or non-stationary scenarios. The framework integrates MLLMs for GUI interaction and aims to enhance autonomous navigation of modern information systems.
Key facts
- Framework uses a two-phase pipeline: offline discovery and inference.
- Offline phase builds a topological knowledge base from execution logs.
- Inference phase employs Adaptive RAG over a pre-established graph.
- Addresses fragmentation in current task sequence modeling.
- Integrates MLLMs for GUI interaction.
- Targets automatic workflow execution in complex information systems.
- Proposed to improve agent effectiveness in novel scenarios.
- Published on arXiv with ID 2605.28607.
Entities
Institutions
- arXiv