Logos: An Agent Harness on a Cross-Process Bus

The premise that capabilities share one process and failure domain is incorrect for Logos. While the underlying spatiotemporal-composability calculus formalizes dynamic composition where components share a single context, Logos is specifically designed to extend this into a distributed, cross-process architecture.

In the Logos system: Each plugin is an independent operating-system process, eliminating the single point of failure inherent in single-process implementations. Failure isolation is achieved because a fault in one process terminates only that specific component, whereas in a single-process model, a crash suspends every component and session. * Shared state is decoupled from processes and resides in an append-only transcript on the bus, allowing any new process to rebuild a session from the transcript via "cold switching" after a failure.

Logos proves that the reversibility guarantees of the calculus hold across processes by leveraging the statelessness of language-model inference and localizing recovery data within each independent plugin process.

Logos: An Agent Harness on a Cross-Process Bus presents a formal and architectural treatment of dynamic agent assembly, framing agent composition as a problem in a spatiotemporal-composability calculus rather than as an ad hoc orchestration pattern. The central object is an agent harness that operates over a cross-process bus, allowing agents or capability modules to be composed, invoked, and coordinated across process boundaries while preserving explicit spatial and temporal structure. In this model, capabilities are treated as tracked-inverse plugin components: they are scoped, attachable units of authority or functionality whose effects, dependencies, and lifecycle can be tracked, and whose inverse or rollback semantics are part of the formal account. This positions agent composition as a disciplined, capability-aware process rather than a loose collection of prompts, tools, and message handlers.

The key contribution is the combination of a formal calculus with a concrete harness design. The calculus makes explicit how agent components can be assembled over time and space, what capabilities they require, and how those capabilities interact under composition. The “tracked-inverse” property is especially important: by associating capabilities with tracked inverses, the system can reason about revocation, cleanup, accountability, and safe detachment. At the same time, the brief’s emphasis on components sharing one process and failure domain suggests a design tradeoff: the bus provides cross-process coordination, while the core capability plugins remain co-located in a single process and failure domain, reducing cross-memory hazards and making local reasoning about state, failure, and recovery more tractable. The result is a model in which dynamic assembly is expressive enough for multi-agent systems but constrained enough to support correctness, isolation, and auditability.

This matters because current agent frameworks often prioritize flexibility over compositional guarantees, leading to systems that are difficult to reason about, secure, or debug. By formalizing agent assembly around capabilities, temporal composition, and a cross-process bus, the work offers a foundation for building agent platforms where dynamic tool use, delegation, and revocation are first-class concerns. For technically sophisticated readers, the paper is significant not merely as an engineering harness, but as an attempt to give agent systems a more rigorous compositional semantics—one that can support safer deployment, clearer failure boundaries, and more predictable behavior in distributed or multi-agent settings.