Logos: An Agent Harness on a Cross-Process Bus
Logos is a cross-process agent harness where a plugin is a process and the only shared state is an append-only transcript. It is built on four lemmas derived from the spatiotemporal-composability calculus and the statelessness of language-model inference. In testing, eighty sessions resumed with no repeated effect after kills placed at the four boundaries of the tool-call cycle.
The paper introduces Logos, a ROS-like cross-process agent harness. It argues that the spatiotemporal-composability calculus and language-model statelessness do not require an agent to run in a single process. Logos implements plugins as separate processes with an append-only transcript as the sole shared state. The system demonstrated fault tolerance by resuming eighty sessions without repeated effects after process kills at the four boundaries of the tool-call cycle.
Logos separates agent components into processes, reducing the blast radius of failures. The append-only transcript serves as the shared state, enabling deterministic recovery. The four lemmas formalize why cross-process composition preserves the soundness invariant of the calculus. The kill-boundary tests suggest the harness can handle process death at critical points in the tool-call cycle without duplicating side effects.
This approach addresses a practical limitation of single-process agent frameworks: a fault in one plugin can suspend all components and interrupt sessions. Cross-process isolation could improve reliability for production agent deployments. The ROS-like design may appeal to robotics and industrial systems where process isolation is already common.
For organizations running long-lived or mission-critical agents, cross-process isolation can reduce downtime and simplify fault recovery. The append-only transcript may also support auditing and replay. However, the paper does not provide commercial deployment data, so business value remains prospective.
Observable next signals include whether Logos is released as open-source software, whether it is adopted by existing agent frameworks, and whether follow-up work benchmarks its overhead against single-process harnesses. Additional evidence may show integration with specific model providers or enterprise platforms.