arXiv
OpenRath: Session-Centered Runtime State for Agent Systems
Fukang Wen, Zhijie Wang, Ruilin Xu
Make Everything Agentic
Rath Lab is an AI research lab at Shanghai Lynwe Technology. We study the core capabilities of agents and harnesses.
Agent-centric, not model-centric
| Dimension | Model-centric | Agent-centric |
|---|---|---|
| Primary output | Models | Agents and harnesses |
| Relation to a domain | Model plus domain, adapted by fine-tuning | Agent plus domain, adapted by redesigning the harness |
| The central question | Where is the ceiling of model capability | Given today’s models, how do you organize them into agents that are reliable over long horizons |