Sign InOpen Brain
VercelEngineering PostOfficial Source

TypeSafe AI's Jev now available on AI Gateway

Jev gives agent loops typed decisions and probabilities without generating prose, offering a smaller control primitive for routing tools, retries, reviews, and guardrails.

Vercel · Sep 16, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Jev accepts shared state and evaluates named questions in parallel, returning typed **Choice, Score, and Boolean** answers with probabilities. It is available through the experimental evaluate API in **AI SDK 7.0.105+**.

Practical Implication

Use it for narrow control-plane decisions in agent loops: tool or subagent selection, retry versus stop, risk scoring, and escalation to human review. Preserve question IDs and calibrate its confidence against labeled workflow examples.

Agent-Ready Context
Jev accepts shared state and evaluates named questions in parallel, returning typed **Choice, Score, and Boolean** answers with probabilities. It is available through the experimental evaluate API in **AI SDK 7.0.105+**.

Use it for narrow control-plane decisions in agent loops: tool or subagent selection, retry versus stop, risk scoring, and escalation to human review. Preserve question IDs and calibrate its confidence against labeled workflow examples.

The reported **193.6x speed** and **444.6x cost** advantages come from TypeSafe's own workflow evaluations, so they do not establish performance on your workload. The API is also experimental.
Connected Context · Feed7 Judgment

Jev narrows model use inside agent harnesses to fast, typed control-plane judgments rather than open-ended execution. Its probabilities can support routing, stopping, and escalation, but require stable question IDs and calibration on labeled workflow data. The experimental API and vendor-reported advantages make it an evaluator to test alongside deterministic safeguards, not evidence that existing reliability controls can be removed.

Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, MaerskJev could encode narrow routing or escalation decisions within the production harness described by Maersk, but it does not replace executable SOPs, bounded tools, replay, or expert correction.AI Agents Are Just Distributed Systems Now — Salman Munaf, TikTokA typed retry-or-stop judgment can govern an agent loop, while the distributed-systems requirements for idempotency, reconciliation, and recovery still determine whether the resulting action is safe.Stop AI Agent Hallucinations: 5 Techniques + Production Patterns - Elizabeth Fuentes, AWSJev supplies a probabilistic decision component that can complement preflight checks and enforced permissions, but the candidate’s external verification and action records remain necessary around irreversible work.Twin: Playing an Unknown Game with a Test-Time Digital TwinTwin gates actions through executable transition validation, providing a stronger domain-specific check than Jev’s typed probabilistic answers where cheap replay and simulation are available.
Context Map
agentcoding#harness-engineering#tool-use#agent-reliability
Uncertainty
The reported **193.6x speed** and **444.6x cost** advantages come from TypeSafe's own workflow evaluations, so they do not establish performance on your workload. The API is also experimental.