Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Juejin test: browser agents read the web better than they write to it

AI summary

In a first-person account of running a browser agent through Juejin's daily routine tasks, the author finds that reading pages works well while write actions such as liking and commenting leak: the tool layer can return success while no backend request was sent, so the agent records a like that never happened. The author recommends giving every write action a postcondition (checking for the network request or the UI text changing from “like” to “liked”), stopping on the first miss instead of re-clicking, cooling down and lowering frequency when the platform's risk control intercepts the action, and ending the run and handing back to a human when a CAPTCHA appears. Because a session was auto-compressed mid-run, the author also prescribes a four-part handover summary (completed items with evidence, unfinished items, explicit prohibitions, next-step order) so constraints are not lost. The article groups browser-agent approaches into vision-based control, model-native action spaces, multi-tab operation with human-in-the-loop, and structured DOM or accessibility-tree interaction, and suggests caching-friendly snapshots. The specific agent and model used are not identified.

In-depth insight

The account converts a single run into a concrete engineering checklist: verify state changes after writes, stop rather than retry blindly, preserve constraints and completed items across context compression, and hand back sensitive actions such as publishing, payment, security confirmation and CAPTCHAs. The author argues the practical ceiling of these agents is set by verification rather than by model reasoning, though the findings are one person's observations without independent measurement.

前一天我让一个能自己操作浏览器的 Agent 去跑掘金的每日任务,本来只是想看看它对真实站点的适配能力,结果掉进了工程细节里:真正难的不是让模型「知道该点什么」,而是让整个循环在真实前端面前不出错。这篇文章不谈模型榜单,只谈我在这条链路上遇到的几个工程真问题,以及背后各家的取舍。

一、四条路线,同一个假设

目前主流的浏览器 Agent 大致分成四类:

  • 视觉为主:Anthropic 的 computer use 属于这一类,给模型截图,输出鼠标坐标与键盘动作。优点是泛化——任何界面都能试;代价是一屏 1080p 截图动辄上千 token,坐标回归差几十像素就点空。
  • 模型内置动作空间:OpenAI 的 Operator / CUA 类方案,把「浏览与点击」当成模型的原生动作,卖点是不需要站点配合。
  • 多标签与人在环:Google 的 Project Mariner 取向,强调并行标签页与关键动作的人工确认,把不确定性包到交互设计里。
  • 结构化 DOM / 可访问性树:国内 IDE 类工具(包括字节 Trae 的 browser 能力)与大量 MCP 服务的选择:把页面上可交互的元素规整成带 ref 的列表,模型只需要说「点 e42」。

#Anthropic computer use#OpenAI Operator#OpenAI CUA#Google Project Mariner#Trae#Cursor#MCP#Juejin

Original · 5 min readRead the original
Group chat

No comments yet. Start the conversation.