Fukang Wen and researchers Zhijie Wang and Ruilin Xu introduced OpenRath on June 17, a programming framework for multi-agent systems that centralizes runtime state management. The system utilizes a 'Session' as a first-class runtime abstraction, allowing developers to handle complex agent interactions using a programming model inspired by PyTorch. By carrying conversation data, tool usage records, memory events, and workspace placements directly within the Session object, the platform eliminates the need to reconstruct state from fragmented external logs.

This design enables explicit fork, merge, and replay operations as native runtime commands. According to the authors, the Session object functions as the primary value passed between agents and workflows, incorporating lineage metadata and token usage. The framework further defines core components including Sandbox, Tool, Agent, Memory, Workflow, and Selector, with the latter used to route control flow decisions during execution. The researchers claim this approach provides a reliable method for auditable composition in agentic systems, though they noted that broader quantitative comparisons remain for future evaluation.

Fukang Wen and researchers Zhijie Wang and Ruilin Xu introduced OpenRath on June 17, a programming framework for multi-agent systems that centralizes runtime state management. The system utilizes a 'Session' as a first-class runtime abstraction, allowing developers to handle complex agent interactions using a programming model inspired by PyTorch. By carrying conversation data, tool usage records, memory events, and workspace placements directly within the Session object, the platform eliminates the need to reconstruct state from fragmented external logs.

This design enables explicit fork, merge, and replay operations as native runtime commands. According to the authors, the Session object functions as the primary value passed between agents and workflows, incorporating lineage metadata and token usage. The framework further defines core components including Sandbox, Tool, Agent, Memory, Workflow, and Selector, with the latter used to route control flow decisions during execution. The researchers claim this approach provides a reliable method for auditable composition in agentic systems, though they noted that broader quantitative comparisons remain for future evaluation.