Logos: A ROS-like Agent Harness Running on a Cross-Process Bus

Forum topic · 小凯 · 2026-09-01

Summary

Logos is a cross-process agent framework presented in arXiv paper 2608.28553, addressing the fault-domain problem of single-process agent systems. Modern agent systems assemble capabilities at runtime as plugins under the spatiotemporal-composability calculus, where a capability is a component carrying a tracked inverse. Typically these plugins share one process context, so all components sit in a single physical failure domain: one fault suspends every component simultaneously, and process death interrupts every hosted session. The authors observe that neither the model nor the calculus binds an agent to one process; language model statelessness keeps cross-step state outside the model, and the soundness invariant is defined purely on the state space. These observations yield four lemmas, on which Logos is built: a ROS-like framework where each plugin is its own process and the only shared state is an append-only transcript. By placing failure handling at four boundaries of the tool-call cycle, the system recovered eighty sessions without duplicated effects. Fault-injection comparisons against a single-process reference configuration showed that a single fault interrupted every co-resident session there, while in the peer-process construction a fault terminated only one node.

Overview

Paper: Logos: An Agent Harness on a Cross-Process Bus (arXiv:2608.28553)
Field: Machine Learning
Authors: Hanzhang Jia, Liheng Zeng, Hao Cheng, Yi Gao, Bo Ma
Posted: 2026-08-28

Key Points

Modern agent systems assemble capabilities at runtime, a form of dynamic composition formally captured by the spatiotemporal-composability calculus, in which a capability is a component carrying a tracked inverse and agents are assembled as plugins.

The standard plugin form is carried by a single process sharing one context, which places all components in one physical failure domain:

A single fault suspends every component at once.
Process death interrupts every session the process hosts.

The paper's central argument:

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.
The soundness invariant is defined on the state space alone.

These observations condense into four lemmas, whose premises are the calculus assumptions and the statelessness of language-model inference.

The Logos Framework

Built on these lemmas, Logos is a ROS-like cross-process agent framework:

Each plugin runs as its own process.
The only shared state is an append-only transcript.
Failure handling is placed at the four boundaries of the tool-call cycle.

Experimental Results

80 sessions recovered after process termination without duplicated effects.
Fault-injection comparison with a single-process reference configuration: one fault interrupted every co-resident session in the single-process setup, whereas in the peer-process construction a fault terminated only one node.

Abstract (translated from source)

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.

--- *Auto-collected on 2026-09-01.*