A new research paper details an agent harness named Logos, designed to allow artificial intelligence agents to operate across multiple processes. This cross-process capability aims to enhance the resilience and continuity of AI agent sessions, particularly in the face of system failures.
Traditionally, agent systems have composed capabilities dynamically at runtime. This composition often occurs within a single process, meaning a fault in that process can halt all components and interrupt ongoing sessions. The spatiotemporal-composability calculus provides a formal framework for such dynamic composition, where capabilities are treated as components with tracked inverses, and agents are assembled as plugins. However, this plugin form typically resides within a single process, creating a single point of failure.
The Logos harness, as described in the paper "Logos: An Agent Harness on a Cross-Process Bus," challenges this limitation. The researchers posit that neither the modeling nor the calculus inherently binds an agent to a single process. They highlight the stateless nature of language models, which keeps cross-step state external to the model itself. The soundness invariant, they argue, is defined on the state space alone.
Based on these observations, Logos is constructed as a "ROS-like" cross-process harness. In this architecture, each plugin is treated as an independent process. The only shared state between these processes is an append-only transcript, which serves as a persistent record of session steps. This design ensures that if one process fails, other components can continue to operate, and the session can potentially resume from the last recorded state.
The paper demonstrates the system's resilience through testing. Eighty sessions reportedly resumed without repeated effects after simulated failures at critical boundaries of the tool-call cycle. A direct comparison with a single-process configuration showed that a single fault interrupted all co-resident sessions. In contrast, with Logos's peer-process construction, a fault was contained to a single node.
This approach draws on concepts from formal theories of composability. Research in this area has explored temporal composability, which involves reversing a component's side effects upon removal, and spatial composability, related to managing inter-component dependencies. The Logos system appears to build upon these ideas by formalizing revertible effects and reactive coeffects, enabling components (agents) to operate independently while maintaining a shared, albeit limited, state for recovery.
The development of agent harnesses is a growing area in AI research. These harnesses provide the necessary infrastructure around large language models (LLMs) to enable them to perform actions beyond simple response generation. They manage aspects such as prompt construction, state management, tool invocation, and execution coordination. The challenge in evolving these harnesses lies in their complexity, tight coupling, and the behavioral distribution of their components.
The Logos harness's cross-process design addresses a fundamental challenge in agent systems: fault tolerance. By distributing agent components across multiple processes, the failure of one component does not necessarily cascade to the entire system. The use of an append-only transcript as the shared state mechanism provides a persistent log that can facilitate session recovery, a critical feature for long-running or complex agent tasks. This work contributes to the broader effort of creating more robust and reliable AI agent architectures.