You shouldn't use Jev for coding agents and routing (weaveos.com)

The recent insights on the Jev model highlight its limitations for coding agents and routing tasks. While Jev offers a unique approach by taking in the state of a session and returning typed values with probabilities, it may not significantly enhance coding-agent traffic handling. Research at Weave shows that effective model routing predominantly relies on understanding the session's history, and Jev does not provide the necessary state-building capabilities essential for improved routing decisions. The significance of these findings underscores the importance of leveraging historical context when routing tasks, rather than solely depending on model outputs like Jev's. Weave's experiments reveal that enhancements in routing accuracy stem from incorporating prior interactions and contextual features, rather than simply swapping models. Consequently, while Jev could serve as a cost-effective classifier in certain contexts, those aiming to optimize coding agent performance should focus on building a robust state representation and logging detailed session trajectories to guide routing effectively.