Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Holos: A Web-Scale LLM-Based Multi-Agent System for the Agentic Web

Xiaohang Nie, Zihan Guo, Zicai Cui, Jiachi Yang, Zeyi Chen, Leheyi De, Yu Zhang, Junwei Liao, Bo Huang, Yingxuan Yang, Zhi Han, Zimian Peng, Linyao Chen, Wenzheng Tom Tang, Zongkai Liu, Tao Zhou, Botao Amber Hu, Shuyang Tang, Jianghao Lin, Weiwen Liu, Muning Wen, Yuanjian Zhou, Weinan Zhang

Abstract

As large language models (LLM)-driven agents transition from isolated task solvers to persistent digital entities, the emergence of the Agentic Web, an ecosystem where heterogeneous agents autonomously interact and co-evolve, marks a pivotal shift toward Artificial General Intelligence (AGI). However, LLM-based multi-agent systems (LaMAS) are hindered by open-world issues such as scaling friction, coordination breakdown, and value dissipation. To address these challenges, we introduce Holos, a web-scale LaMAS architected for long-term ecological persistence. Holos adopts a five-layer architecture, with core modules primarily featuring the Nuwa engine for high-efficiency agent generation and hosting, a market-driven Orchestrator for resilient coordination, and an endogenous value cycle to achieve incentive compatibility. By bridging the gap between micro-level collaboration and macro-scale emergence, Holos hopes to lay the foundation for the next generation of the self-organizing and continuously evolving Agentic Web. We have publicly released Holos (accessible at https://holosai.io), providing a resource for the community and a testbed for future research in large-scale agentic ecosystems.

Holos: A Web-Scale LLM-Based Multi-Agent System for the Agentic Web

Abstract

As large language models (LLM)-driven agents transition from isolated task solvers to persistent digital entities, the emergence of the Agentic Web, an ecosystem where heterogeneous agents autonomously interact and co-evolve, marks a pivotal shift toward Artificial General Intelligence (AGI). However, LLM-based multi-agent systems (LaMAS) are hindered by open-world issues such as scaling friction, coordination breakdown, and value dissipation. To address these challenges, we introduce Holos, a web-scale LaMAS architected for long-term ecological persistence. Holos adopts a five-layer architecture, with core modules primarily featuring the Nuwa engine for high-efficiency agent generation and hosting, a market-driven Orchestrator for resilient coordination, and an endogenous value cycle to achieve incentive compatibility. By bridging the gap between micro-level collaboration and macro-scale emergence, Holos hopes to lay the foundation for the next generation of the self-organizing and continuously evolving Agentic Web. We have publicly released Holos (accessible at https://holosai.io), providing a resource for the community and a testbed for future research in large-scale agentic ecosystems.

Paper Structure

This paper contains 67 sections, 4 equations, 8 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Motivation and Overview
  2. Problem Setting and Challenges
  3. System Design of Holos
  4. Design Principles
  5. Five-Layer Architecture
  6. Interface Layer
  7. Decentralized Discovery and Identity Standardization
  8. Secure Interoperability and Context Propagation
  9. Substrate Layer
  10. The Anatomy of an Agent
  11. Evolutionary Synthesis of Agent Population
  12. Dynamic Instantiation and Runtime Environment
  13. Coordination Layer
  14. Intent Orchestration with Topological Validation
  15. Formal Modeling of Abstract Task DAG.
  16. ...and 52 more sections

Figures (8)

  • Figure 1: Screenshot of Holos' user interface. a is the user entry on the homepage, where users can submit requests via natural language; b is the directed acyclic graph (DAG) of the Orchestrator during task execution; c shows details of a subtask, including mission responsibility, results, and evaluation; d displays agent outputs and an interactive workspace; e visualizes all agents as a galaxy and highlights selected agents with detailed information.
  • Figure 2: The five-layer architecture of Holos. This architecture orchestrates the end-to-end lifecycle from genetic agent synthesis to market-driven task execution. Through resilient coordination and evolutionary feedback loops, it fosters a synergistic environment for continuous co-creation and collective intelligence growth.
  • Figure 3: Results of the scale-invariant discovery efficiency test. The left subgraph displays the skills of the inserted "'Needle' agent and the query issued during the experiment. The right subgraph illustrates the search steps and elapsed time required to locate a specific agent with extremely sparse semantic features across agent populations spanning several orders of magnitude.
  • Figure 4: Results for adaptive self-healing resilience under varying failure injection probabilities ($P$). The left subgraph illustrates the execution efficiency (mean steps) and task reliability (success rate), showing how the system maintains resilience despite increased failure rates. The right subgraph evaluates the self-healing overhead by comparing the total errors encountered against the mean attempts required for successful task, highlighting the responsiveness to escalating stress.
  • Figure 5: Analysis of resource allocation effectiveness and ability identification efficiency of Holos. The subgraph on the left shows the trend of the Spearman correlation coefficient between the ability of agents and their revenue as them evolve over time. The subgraph on the right displays the statistical distribution of evaluation scores for agents with different ability of the last 10 epochs.
  • ...and 3 more figures