Overview & result
eve is Vercel's filesystem-first framework for durable AI agents. Jev is not mentioned in the main README; the framework's docs say its auto helper chooses an agent model from an allowlist before inference and by default evaluates with typesafe-ai/jev through Vercel AI Gateway. An evaluate function lets tools and application code ask typed questions the same way.
From the original post
The Open Framework for Building Agents
How Jev fits in the loop
- An eve agent declares its model with auto() and an allowlist of models, each with a short description.
- Before inference, the evaluator chooses a model from the allowlist; the default evaluator is typesafe-ai/jev.
- The request goes through Vercel AI Gateway; eve adds no TypeSafe credential or transport layer.
- Tools and application code can ask their own typed questions with evaluate.
How to reproduce
- Inspect the original showcase and source material at https://github.com/vercel/eve
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
A mainstream agent framework uses Jev as its default for per-turn model selection.
Limitations
- Jev is not mentioned in the main README; the details come from the framework's docs.
- The docs say the AI SDK evaluation model specification is experimental and can change in patch releases.
- Gateway authentication is required.