Skip to content

OxyGent: Making Multi-Agent Systems Modular, Observable, and Evolvable via Oxy Abstraction

🕒 Published (v1): 2026-04-28 13:08 UTC · Source: Arxiv · Venue: ACL 2026 · link

Why this paper was selected

Production multi-agent framework with observability, modularity, evolvability; open-source

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

OxyGent is an open-source multi-agent system (MAS) framework from JD.com that introduces a unified "Oxy" abstraction treating agents, tools, LLMs, and reasoning flows as interchangeable atomic nodes, combined with an AI asset management platform (OxyBank) for continuous self-improvement. It replaces static DAG workflows with permission-driven dynamic planning that synthesizes execution graphs at runtime. On GAIA (July 2025), it scored 59.14%, ranking second among open-source methods.

Problem

Production-ready MAS deployment is hindered by three critical gaps: (1) lack of standardized abstractions across agents, tools, and flows, causing repetitive implementation and low reusability; (2) rigid plan-and-execute workflows that break under dynamic, uncertain environments; and (3) absence of closed-loop evolution mechanisms, causing agent performance to stagnate or degrade after deployment.

Method

Unified Oxy Abstraction. Every AI primitive—agents, tools, LLMs, reasoning flows—is wrapped as a pluggable Oxy node with a consistent interface, enabling hot-swapping at runtime without refactoring. State management uses a four-tier hierarchical data scoping system: Application (global), Session Group (cluster-shared), Request (trajectory-scoped), and Node (local), enforced via typed accessor APIs.

Permission-Driven Dynamic Planning. Rather than predefined DAGs, collaboration topologies emerge from authorization relationships between Oxy nodes. At runtime, the system traverses a standardized execution lifecycle per node (pre-process → execute → post-process → format output) and synthesizes real-time call-graph visualizations. Aspect-Oriented Programming (AOP) injects cross-cutting concerns (security auditing, performance monitoring) at lifecycle joinpoints without polluting business logic.

OxyBank Evolution Engine. A closed-loop AI asset platform captures online execution traces, passes them through MD5-based deduplication and priority inference (call-chain metadata determines P0/P1/etc.), enforces a strict state machine (pending → annotated → approved) before any sample enters the knowledge base, and feeds approved samples back to update knowledge bases or fine-tune models. An AI-driven Optimize Prompt module extracts best practices from validated traces to iteratively refine agent system prompts.

Key Contributions

  • Unified Oxy abstraction: single interface for all AI primitives enabling Lego-like, hot-swappable MAS composition
  • Permission-driven dynamic planning: runtime-synthesized execution graphs replace static DAGs, with real-time AOP-based visualization
  • OxyBank: industrial-grade AI asset management platform with multi-layered quality gating for continuous agent evolution via data backflow
  • Production observability: per-agent time tracking, bottleneck identification, interactive decision exploration with rollback and parallel variant comparison
  • Empirical validation: GAIA score 59.14% (2nd among open-source, July 2025); industrial deployment at 2,000+ agent scale improving e-commerce classification from 61.3% → 85.6%
  • Full open-source release under Apache 2.0

Results

  • GAIA benchmark: 59.14% average (Level 1: 77.42%, Level 2: 56.60%, Level 3: 32.65%), second among open-source methods behind OWL++ (60.8%) as of July 22, 2025
  • Ablation on GAIA (sequential additions over single-agent baseline of 36.21%):
  • +Multi-agent: 42.19%
  • +Dynamic planning: 52.16% (+18.24% on Level 2 specifically)
  • +Memory backflow: 59.14% (+15.05% on Level 1)
  • E-commerce classification (2,000+ agent hierarchical MAS): 85.6% vs. 61.3% RAG+DeepSeek-R1-Distill-Qwen-32B baseline; autonomously discovers ~5.4 new categories/week; inference latency 2.3Ă— higher than single-agent baseline

Limitations

  • Large-scale training in OxyBank still requires manual resource configuration; intelligent resource scheduling for a fully automated lifecycle is in progress
  • Ambiguous semantic queries can induce repetitive ReAct loops or hallucinations, currently mitigated only by trace replay and manual auditing
  • Latency overhead of 2.3Ă— over single-agent baselines limits suitability for real-time (online) environments

Relevance to Agentic AI / LLM Agents

OxyGent directly addresses the core engineering challenges of production MAS deployment: compositional modularity, runtime observability, and self-improvement—three properties that most academic agent frameworks treat separately or ignore. The permission-driven dynamic planning model is a concrete alternative to the dominant static-DAG paradigm (LangGraph, CrewAI), offering a path toward truly adaptive agent topologies without sacrificing inspectability. The OxyBank closed-loop evolution engine operationalizes the growing research interest in self-evolving agents by anchoring it to industrial-grade data quality controls, making it directly relevant to work on agent continual learning and data-flywheel approaches. The GAIA benchmark positioning and the 2,000-agent industrial case study provide credible evidence that the framework's design choices transfer beyond toy settings.