Jev as the routing layer for multi-harness agents
The author argues routing to Claude Code, Codex, or Opencode should move out of system prompts and into Jev, which returns only a typed decision with a confidence score, claiming it is 193x faster and 444x cheaper.
The routing problem in a multi-harness agentic stack has been an LLM call problem. You route to Claude Code, Codex, Opencode, or another harness by putting logic in a system prompt and burning tokens on every decision. Jev changes that architecture. TypeSafe AI, led by Diogo Almeida who helped ship the original ChatGPT, released it last week. It's not a chat model. It returns only a typed decision with a confidence score. No output tokens billed. 193x faster than Claude Sonnet 5 on classification tasks. 444x cheaper than Opus 5 on the same workload. The immediate use case is routing: give Jev the task description and a list of available harnesses, and it returns a decision in milliseconds at near-zero cost. That's what the demo above is doing: a Claude Code plugin that reads the local system's available harnesses and uses Jev to pick the right one for each task. The broader pattern I keep noticing: decisions and generation are separating into different model classes. You don't need a frontier LLM to classify a task, route a request, or run a quality gate. You need a model trained specifically for structured output under time and cost pressure. The products converging here: TypeSafe's Jev for routing. Groq and Cerebras on the inference side for latency. Braintrust, Arize, and PromptLayer building the eval harnesses that feed signal back into routing logic. The agentic stack is splitting into two layers. LLMs for generation. Decision models for control flow. https://x.com/mdlahfir/status/2100399709995356266
This post quotes another post: “Claude Code plugin uses Jev to pick the best local model for a task”