Use case
Routing is currently done by an LLM (Bedrock, Anthropic or OpenAI classifier): a full model call per request, whose free-text or tool-call output has to be parsed into an agent name.
Solution/User Experience
Add a JevClassifier, in both TypeScript and Python, backed by TypeSafe's Jev System One model (docs). Jev returns a typed decision: the registered agents are sent as the options of a choice question, and Jev returns the chosen agent plus a calibrated confidence. There's no output parsing, and it costs $0.042 per million input tokens with output free.
new AgentSquad({ classifier: new JevClassifier() }) // TYPESAFE_API_KEY setAgentSquad(classifier=JevClassifier())
Scope: the classifier in both runtimes with matching options and tests, a docs page (including limitations, linked to the official docs), and demos in both languages (multi-turn routing, plus a Jev vs Bedrock classifier comparison of accuracy, latency and cost).
Alternative solutions
Keep using an LLM classifier. On a 12-turn test conversation with interleaved follow-ups, Jev got every turn right at a fraction of the cost and latency of BedrockClassifier on Claude Opus 5.
Use case
Routing is currently done by an LLM (Bedrock, Anthropic or OpenAI classifier): a full model call per request, whose free-text or tool-call output has to be parsed into an agent name.
Solution/User Experience
Add a
JevClassifier, in both TypeScript and Python, backed by TypeSafe's Jev System One model (docs). Jev returns a typed decision: the registered agents are sent as the options of achoicequestion, and Jev returns the chosen agent plus a calibrated confidence. There's no output parsing, and it costs $0.042 per million input tokens with output free.Scope: the classifier in both runtimes with matching options and tests, a docs page (including limitations, linked to the official docs), and demos in both languages (multi-turn routing, plus a Jev vs Bedrock classifier comparison of accuracy, latency and cost).
Alternative solutions
Keep using an LLM classifier. On a 12-turn test conversation with interleaved follow-ups, Jev got every turn right at a fraction of the cost and latency of
BedrockClassifieron Claude Opus 5.