Overview
This research introduces Logos, a cross-process agent harness designed to support modern agent systems. Logos implements a structure where each agent plugin operates as an independent process, contrasting with traditional single-process agent constructions. The system utilizes an append-only transcript for shared state among these processes.
The development of Logos is grounded in observations derived from the spatiotemporal-composability calculus and the statelessness of language-model inference. These observations condense into four lemmas, whose premises are the hypotheses of the calculus and the mentioned statelessness. The harness aims to address limitations associated with agents being confined to a single process, particularly concerning fault domains and session continuity.
Research Context
Modern agent systems are characterized by their ability to dynamically assemble capabilities at runtime. This dynamic composition has received a formal treatment through the spatiotemporal-composability calculus. In this calculus, a capability is defined as a component that carries a tracked inverse, and agents are assembled as plugins. Historically, such plugins have been implemented within a single process, sharing a single context. This single-process carrier implies that all components reside within a common physical failure domain. Consequently, a fault in one component can suspend every co-resident component, and the termination of the process interrupts all sessions hosted by that process.
The current work posits that neither the spatiotemporal-composability calculus's modeling nor its formal definitions necessitate binding an agent to a single process. It emphasizes that the statelessness inherent in the language model keeps all cross-step state external to the model itself. Furthermore, the soundness invariant for such systems is defined solely on the state space. These foundational aspects informed the design principles for Logos.
Approach
The development of Logos proceeds from four specific lemmas. These lemmas derive their premises from the hypotheses embedded within the spatiotemporal-composability calculus, coupled with the principle of statelessness in language-model inference. Based on these theoretical foundations, Logos was constructed as a ROS-like cross-process agent harness. In this architecture, a key distinction is that each plugin functions as an individual process. The only shared state across these distinct processes is maintained through an append-only transcript.
To evaluate its resilience and fault tolerance, Logos was subjected to testing involving the placement of kills at four distinct boundaries within the tool-call cycle. The system's performance was then compared against a reference configuration consisting of a single-process agent system.
Findings
The conducted experiments demonstrated that Logos enabled 80 sessions to resume without exhibiting any repeated effect after kills were introduced. This indicates a robust recovery mechanism. In a direct same-fault comparison with a single-process reference configuration, the single-process setup showed that one fault invariably interrupted every co-resident session. In contrast, under the peer-process construction implemented by Logos, a single fault was contained and ended at only one node. This isolation suggests a significant improvement in fault tolerance and system stability compared to the single-process model.
Why This Matters
The findings regarding Logos indicate a method for constructing agent systems with enhanced fault isolation. By allowing each plugin to exist as a separate process and limiting shared state to an append-only transcript, the system mitigates the risk of a single fault propagating across all active sessions. This directly addresses the vulnerability of single-process agent architectures where a fault in one component can disrupt all co-resident components and terminate all hosted sessions. The observed ability of 80 sessions to resume without repeated effects after targeted kills suggests improved resilience for dynamic agent systems operating under the spatiotemporal-composability calculus.