Hello everyone, I am PaperAgent, not just an Agent!
In an era where large language model capabilities are leaping forward, Multi-Agent Systems (MAS) have become a key pillar for tackling highly challenging tasks. However, examining the current MAS orchestration landscape, developers remain trapped by outdated construction methods: either investing high engineering costs to manually maintain complex communication logic via hard-coding, or compromising with drag-and-drop canvas frameworks that involve heavy workloads for developing complex multi-agent systems and lack the ability to integrate AI to replace human labor like Vibe Coding.
To break this inefficient orchestration paradigm, Beijing University of Posts and Telecommunications has officially open-sourced a new framework: MASFactory. This framework proposes a Graph-Centric architecture to describe multi-agent workflows and introduces the "Vibe Graphing" development paradigm, pushing MAS development from manual assembly into the era of natural language drive.
Vibe Graphing
From "Hard-Coding" to "Intent-Driven" Vibe Graphing. Unlike traditional node-linking or writing underlying logic, MASFactory advocates "global intent first, local details later." Developers only need to articulate the system's ultimate goals and expected role divisions in natural language; the built-in AI engine will then rapidly deduce a feasible collaborative graph structure.
To address the issue of uncontrollable hallucinations that may occur during AI generation, MASFactory introduces a "Human-In-the-Loop" process. Developers can review the AI's proposal at the end of each stage and provide modification feedback until the AI's solution satisfies the developer!
Comprehensive Comparison of Multi-Agent Orchestration Frameworks
From the perspective of multi-agent system development evolution, current mainstream methods can be roughly divided into three categories: code writing, visual drag-and-drop, and Vibe Graphing.
Code writing offers the highest flexibility and scalability, suitable for building complex, multi-layered multi-agent collaborative systems, but it demands high developer capability and incurs greater overall development costs.
Visual drag-and-drop significantly lowers the usage threshold, allowing more users to quickly build basic workflows, but it still faces limitations in complex topologies, fine-grained control, and subsequent iterations.
In contrast, Vibe Graphing further advances multi-agent system design from "manual implementation" to "intent-driven generation." Users do not need to write extensive code or repeatedly drag nodes; they only need to clearly describe requirements and continuously refine the design during interaction to quickly complete system construction and iteration. Precisely for this reason, Vibe Graphing is more suitable for rapid prototyping and low-labor-cost development under complex requirements.
MASFactory System Architecture
MASFactory scientifically abstracts complex multi-agent interactions into four layers:
Graph Skeleton Layer: Like other multi-agent workflow frameworks, it uses Nodes and Edges as the underlying foundational skeleton to depict collaboration and message flow between agents.
Component Layer: Encapsulates the foundational skeleton into out-of-the-box modules. This includes not only Agents that execute tasks but also introduces Switches for dynamic routing, Loops for handling multi-round games, and Human nodes for human-in-the-loop scenarios. More importantly, any Graph can be nested infinitely as a sub-node, achieving extreme logic reusability.
Unified Protocol Layer: Leveraging adapter mechanisms, it seamlessly smooths out differences in various communication protocols and uniformly manages context environments, easily integrating enhanced capabilities like RAG and Memory.
Hybrid Interaction Layer: Provides extremely flexible operation entry points for upper-layer applications, including declarative and imperative orchestration layers compatible with code development, visual drag-and-drop orchestration layers, and Vibe Graphing orchestration layers, accommodating the usage habits of different developers.
Performance Comparison
To systematically evaluate the effectiveness of MASFactory, the paper conducted experimental verification from two aspects:
First, examining MASFactory's ability to reproduce existing representative multi-agent systems;
Second, verifying the actual effects of workflows generated by Vibe Graphing. The experiments covered 7 public benchmarks, including HumanEval, MBPP, BigCodeBench, and SRDD for code tasks, as well as MMLU-Pro, GAIA, and GPQA for general reasoning and tool usage.
Overall, the results of MASFactory on the 7 public benchmarks indicate that it can stably carry different types of multi-agent workflows, also proving that the path of "Natural Language Intent — Editable Specification — Executable Graph" is feasible.
At the same time, the workflows generated by Vibe Graphing also demonstrate strong competitiveness. Whether it is Vibe Graphing-ChatDev adapted from ChatDev or Vibe Graphing-Task Specific generated directly for tasks, both achieved considerable results on multiple benchmarks. Among them, the Task Specific scheme performed outstandingly on HumanEval, BigCodeBench, and SRDD, indicating that workflow generation driven by natural language has been able to approach or even reach the effect of manually designed systems.
- **Code Repository:** https://github.com/BUPT-GAMMA/MASFactory
- **Paper Address:** https://arxiv.org/abs/2603.06007Hands-on Design of AI Agents: (Orchestration, Memory, Plugins, Workflow, Collaboration)
Sharing Two Latest Claude Skills Papers with 3 Core Conclusions