跳转至

Logos：基于跨进程总线的智能体运行框架

文章背景与核心概要

现代大语言模型智能体系统通常将动态的运行时功能组装到共享单一上下文的单个进程中。然而，这种单进程模型创造了一个统一的故障域，其中单点故障会导致所有共驻会话同时崩溃。为了解决这一架构痛点，本文推出了 Logos，一个类似于 ROS（机器人操作系统）的跨进程智能体运行框架（Harness）。

通过利用语言模型推理的无状态特性，并将合理性不变式（soundness invariants）仅定义在状态空间上，Logos 将智能体插件解耦为通过单个只追加（append-only）记录总线连接的独立进程。该架构确保了工具调用周期中任何一点的故障都能被隔离在单个节点内，从而允许八十个并发会话无缝恢复且不会产生重复的副作用。这项研究在提升智能体系统的鲁棒性与容错能力方面迈出了重要一步。

Logos: An Agent Harness on a Cross-Process Bus

Logos: An Agent Harness on a Cross-Process Bus

Summary

Modern agent systems often assemble dynamic runtime capabilities into a single process sharing one context. However, this single-process model creates a unified failure domain where a single fault crashes every co-resident session simultaneously.

This paper introduces Logos, a ROS-like cross-process agent harness. By leveraging the statelessness of language model inference and defining sound invariants solely on the state space, Logos decouples agent plugins into independent processes connected by a single append-only transcript bus. This architecture ensures that a fault at any point in the tool-call cycle isolates the failure to a single node, allowing eighty concurrent sessions to seamlessly resume without duplicate effects.

Document Metadata

Field Details
Title Logos: An Agent Harness on a Cross-Process Bus
Authors Hanzhang Jia, Liheng Zeng, Hao Cheng, Yi Gao, Bo Ma
Subjects Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
arXiv Identifier arXiv:2608.28553 [cs.AI]
DOI 10.48550/arXiv.2608.28553
Submission History • v1: Aug 28, 2026
• v2: Aug 31, 2026 (Current)

Abstract

Modern agent systems assemble capabilities at runtime, and this dynamic composition has recently received a complete formal treatment in the spatiotemporal-composability calculus, in which a capability is a component carrying a tracked inverse, and agents are assembled as plugins. This plugin form is carried by a single process sharing one context, a carrier that places all components in one physical failure domain, a fault suspends every component at once, and process death interrupts every session the process hosts.

This paper shows that neither the modeling nor the calculus binds an agent to one process, the statelessness of the language model keeps all cross-step state outside the model, and the soundness invariant is defined on the state space alone. These observations condense into four lemmas whose premises are the hypotheses of the calculus and the statelessness of language-model inference.

On these lemmas this paper constructs Logos, a ROS-like cross-process agent harness in which a plugin is a process and the only shared state is an append-only transcript. Eighty sessions resume with no repeated effect after kills placed at the four boundaries of the tool-call cycle, and a same-fault comparison with a single process reference configuration shows one fault interrupting every co-resident session while under the peer-process construction one fault ends at one node.