@SUOHA_AI: JEV's capability boundaries are slowly starting to reveal themselves. A few days ago, the explosive demo open-sourced b…

X AI KOLs Following News
ai-agent browser-automation performance-comparison reasoning open-source jev luna

Summary

The article compares the AI agent JEV with Luna in complex browser automation tasks, highlighting JEV's speed in simple scenarios but its failure in real-world complexities due to lacking reasoning abilities.

JEV's capability boundaries are slowly starting to reveal themselves. A few days ago, the explosive demo open-sourced by Browser Use founder @gregpr07 (booking a flight in 7 seconds, cost $0.0039) took the entire internet by storm. But then he personally ran a round of long-chain task tests with JEV and found the accuracy rate to be very low, far behind GPT luna: • Luna: 17 / 20 • Jev Agent: 1 / 20 Out of 20 complex real-world tasks, JEV completely failed 19 of them... So why was it able to book a flight in 7 seconds before? Where exactly is it strong? ✅This is what JEV excels at: Input request ──> Fill form ──> Select flight ──> Submit successfully (no branches at all, all straight roads) 💥This is what JEV can't do: 【Real complex long tasks】 Input request ──> Popup blocker? ──> Captcha? ↓ ↓ [Dead end] [Dead end] │ │ JEV: Zero reasoning, no backtracking, direct deadlock on the spot 💥 Luna + reasoning model: Retreat one step (Backtrack), switch strategies to bypass ✅ Objectively speaking, JEV is indeed unbelievably fast, but once it enters a real complex environment, the logic completely collapses: Browser Use founder says: Browser interaction has never been simple element clicking, but an extremely complex【high-dimensional state space search】 In a real webpage, a qualified Agent must possess two core underlying behaviors: Stop and think deeply (Think hard) When hitting a dead end, backtrack the original path (Go back / Backtrack) While JEV's reasoning ability (Reasoning) is completely 0. Looks like Jev still has a long way to go~~
Original Article
View Cached Full Text

Cached at: 09/21/26, 05:27 AM

JEV’s capability boundaries are slowly starting to reveal themselves.

A few days ago, the explosive demo open-sourced by Browser Use founder @gregpr07 (booking a flight in 7 seconds, cost $0.0039) took the entire internet by storm.

But then he personally ran a round of long-chain task tests with JEV and found the accuracy rate to be very low, far behind GPT luna:

• Luna: 17 / 20 • Jev Agent: 1 / 20

Out of 20 complex real-world tasks, JEV completely failed 19 of them…

So why was it able to book a flight in 7 seconds before? Where exactly is it strong?

✅This is what JEV excels at: Input request ──> Fill form ──> Select flight ──> Submit successfully (no branches at all, all straight roads)

💥This is what JEV can’t do: 【Real complex long tasks】 Input request ──> Popup blocker? ──> Captcha? ↓ ↓ [Dead end] [Dead end] │ │ JEV: Zero reasoning, no backtracking, direct deadlock on the spot 💥 Luna + reasoning model: Retreat one step (Backtrack), switch strategies to bypass ✅

Objectively speaking, JEV is indeed unbelievably fast, but once it enters a real complex environment, the logic completely collapses:

Browser Use founder says: Browser interaction has never been simple element clicking, but an extremely complex【high-dimensional state space search】

In a real webpage, a qualified Agent must possess two core underlying behaviors: Stop and think deeply (Think hard) When hitting a dead end, backtrack the original path (Go back / Backtrack) While JEV’s reasoning ability (Reasoning) is completely 0.

Looks like Jev still has a long way to go~~

梭哈.AI (@SUOHA_AI): Browser Use 接上 JEV 这速度太他妈吓人了吧？？

过去大家看 GPT、Claude 操控浏览器，体验其实并不好，因为真的太慢了，中间还经常卡死…

因为让大模型写一堆长篇大论去操作网页，本质上就是杀鸡用牛刀

在网页上点按钮，本质上就是一个离散的多选题，根本不需要 AI 会写诗

JEV 拿到 DOM

Similar Articles