What changed
On 21 September, Vercel announced that AI Gateway can now call Jev from TypeSafe AI through three paths: an existing TypeSafe client, a native HTTP API, or the AI SDK. The integration is more than another model ID. Jev accepts shared application state and typed questions, then returns a choice, score, or Boolean probability instead of generated prose that the application must parse.
Vercel also puts those evaluations in the same gateway billing and observability surface as other model calls. The Jev model page describes it as a decision model for classification, routing, rubric-based assessment, and automated verification, with multiple questions evaluated in parallel.
A complementary ecosystem signal arrived in Pydantic AI v2.46.0: its TypeSafeModel can fill tool arguments and choose between union output types, while a new Choices helper supports options assembled at runtime.
Why it matters
Many agent loops currently ask a general-purpose model to decide whether to continue, retry, escalate, or select a tool, then parse JSON and validate it. That works, but it mixes open-ended generation with a narrow control decision.
Jev makes the decision boundary explicit. A workflow can declare the question, the allowed answer shape, and the threshold that changes the next state. The result is still probabilistic, but the uncertainty is visible to the application instead of being hidden behind a string parser. That makes it easier to route high-confidence cases automatically and reserve ambiguous cases for a stronger model or a person.
A practical pattern
A TypeScript agent can use the AI SDK to ask whether it should take another step:
import { experimental_evaluate as evaluate } from 'ai';
const result = await evaluate({
model: 'typesafe-ai/jev',
state: runState,
questions: {
continueWorking: {
type: 'boolean',
instructions: 'Should the agent take another step?',
},
},
});
const shouldContinue =
result.answers.continueWorking.probability >= 0.8;
The same evaluation can be called from any language through POST /v1/evaluate, or an existing TypeSafe client can be moved behind the TypeSafe-compatible AI Gateway API by changing its base URL and gateway credential.
Takeaway for builders
Use a typed evaluator for small, repeated decisions inside an agent loop, not for open-ended reasoning. Keep the threshold, fallback, and human-review policy in your own code, and log the probability with the eventual outcome. Typed answers remove parsing failure; they do not remove uncertainty.
That is the product signal in this week’s release: AI infrastructure is starting to treat evaluation and control decisions as first-class runtime calls, with their own interface, cost accounting, and operational trail.