OxyGent: Making Multi-Agent Systems Modular, Observ... | AI Research

OxyGent: Making Multi-Agent Systems Modular, Observable, and Evolvable via Oxy Abstraction
Deploying multi-agent systems (MAS) in complex industrial settings often proves difficult because traditional frameworks are frequently rigid, hard to monitor, and difficult to update once deployed. OxyGent is an open-source framework designed to solve these issues by treating agents, tools, and reasoning flows as pluggable, atomic components. By moving away from static workflows toward a modular, "Lego-like" architecture, OxyGent enables developers to build scalable systems that are easier to observe, manage, and continuously improve.

A Unified Modular Architecture

At the heart of OxyGent is the "Oxy" abstraction, which standardizes how different parts of an AI system interact. Instead of treating agents and tools as separate, disconnected layers, the framework encapsulates them into interchangeable nodes. This design is supported by a four-tier data scoping mechanism—Application, Session Group, Request, and Node—which ensures that data is managed efficiently and securely across distributed environments. By using Aspect-Oriented Programming, the framework also separates core business logic from maintenance tasks like security auditing and performance logging, allowing the system to grow without becoming overly complex.

Dynamic Planning and Observability

Traditional MAS frameworks often rely on pre-defined, static workflows that struggle to adapt to unpredictable environments. OxyGent replaces these rigid structures with permission-driven dynamic planning. In this model, the system determines the execution path at runtime based on the authorization relationships between different components. This approach allows the framework to generate real-time, adaptive visualizations of the actual call graphs. Developers can monitor resource consumption, track decision-making trajectories, and even pause or modify intermediate steps during execution, providing deep transparency into how the agents are performing.

Continuous Evolution with OxyBank

To prevent performance stagnation, OxyGent integrates OxyBank, an AI asset management platform that acts as the system's evolutionary engine. OxyBank captures execution traces from the production environment and converts them into high-quality training data. Through a closed-loop process involving automated rewarding and human-in-the-loop annotation, the system can refine its strategies and update its knowledge base over time. This ensures that the collective intelligence of the agents continues to improve through data-centric learning rather than remaining static after deployment.

Performance and Real-World Application

OxyGent has been validated through both academic benchmarks and large-scale industrial use. In evaluations on the GAIA benchmark, the framework demonstrated strong performance in managing long-chain interactions. In a practical business scenario, a hierarchical system built with OxyGent successfully managed over 2,000 agents to perform complex e-commerce classification, significantly improving accuracy compared to single-agent systems. While the framework currently requires some manual configuration for large-scale training, future development is focused on intelligent resource scheduling to create a fully automated lifecycle for agent construction and refinement.