#
On April 3, 2026, a research team from Beijing University of Posts and Telecommunications (BUPT) officially open-sourced its new-generation Multi-Agent System (MAS) orchestration framework—MASFactory. This project introduces an innovative paradigm called “Vibe Graphing,” aiming to systematically address the engineering complexity and high API call costs currently faced in MAS development.
Vibe Graphing: A New Structured Orchestration Paradigm
As the capabilities of Large Language Models (LLMs) advance, solving complex tasks through multi-agent collaboration has become a technical consensus. However, existing orchestration methods have clear limitations. On one hand, hard-coding approaches based on Domain-Specific Languages (DSLs) have high learning costs and are difficult to maintain. On the other hand, purely visual drag-and-drop interfaces fall short when handling complex logic. Even the emerging “Vibe Coding” paradigm suffers from low efficiency and high costs due to LLMs’ insufficient understanding of niche DSLs.
MASFactory’s Vibe Graphing paradigm abstracts the multi-agent construction process into a structured, three-stage compilation flow driven by internal Agent components:
Role Assignment: Upon receiving a natural language command, the system first decomposes the task, mapping it to a set of agent roles with clearly defined responsibilities, such as “document retriever” or “code reviewer,” without delving into implementation details.
Topology Design: After defining the roles, the system constructs a directed graph skeleton that defines inter-node dependencies and execution order (e.g., serial, parallel). This stage focuses on the high-level structure of the workflow, excluding specific prompts or tool call details.
Semantic Completion: Finally, the system performs parameterized filling of the graph skeleton, precisely configuring the required instructions, input/output formats, and other parameters for each node. Since the AI only needs to generate and modify short JSON configurations in this process, rather than lengthy code, token consumption is reportedly reduced exponentially. This can lower API costs to as little as one-tenth of traditional methods for similar tasks.
MASFactory: A Graph-Centric Four-Layer Architecture
The implementation of Vibe Graphing relies on MASFactory’s robust underlying architecture. Centered around a directed computation graph, the system is divided into four logical layers, ensuring its flexibility and scalability.
Graph Skeleton Layer: As the bottom layer, the system decomposes collaboration signals into three independent flows: the Control Flow drives task scheduling and sequencing; the Message Flow is responsible for passing payloads between nodes; and the State Flow synchronizes contextual information within the graph’s hierarchical structure.
Component & Reusable Layer: Above the graph skeleton, the system provides a series of abstract components, including Agent (representing basic intelligent entities), Graph (for subgraph nesting), Loop (for implementing cycles and reflection mechanisms), and Switch (supporting dynamic routing). Through mechanisms like NodeTemplate and ComposedGraph, developers can efficiently reuse topological structures.
Protocol & Context Adapter Layer: To integrate heterogeneous external systems (such as RAG and memory modules), MASFactory introduces the Message Adapter and Context Adapter. The former decouples communication protocols, allowing developers to switch freely between formats like JSON Schema and Markdown; the latter provides standardized interfaces for external frameworks like LlamaIndex and Mem0.
Hybrid Orchestration & Visualization Layer: MASFactory offers extremely high flexibility at the operational level, supporting declarative/imperative code development, visual drag-and-drop, and the Vibe Graphing paradigm. These models can be used in a hybrid fashion within the same project. Additionally, the official VS Code extension, “MASFactory Visualizer,” provides visual support for the entire development, debugging, and execution lifecycle.
Performance Evaluation and Future Prospects
To validate the framework’s effectiveness, the research team conducted extensive evaluations on seven mainstream benchmark datasets, including HumanEval, GAIA, and MMLU-Pro. The results show that MASFactory can not only stably and efficiently replicate well-known industry multi-agent systems like ChatDev, MetaGPT, and AgentVerse but also outperforms the original hard-coded implementations on multiple metrics.
The open-sourcing of MASFactory provides developers of multi-agent applications with a solution that balances low cost, high efficiency, and strong scalability. It is expected to drive the adoption of MAS technology in a wider range of complex scenarios. The project’s code repository and academic paper are now available online for community research and use.
