一个反直觉的观察
让 Agent 操作浏览器完成一组社区日常任务（浏览、点赞、收藏、评论、发布），最容易归因错误的就是失败：任务没完成，第一反应往往是模型不行。真把每一处失败摊开看，模型负责的部分几乎没出问题，出问题的是环境、框架契约和状态确认。
第一类卡点：环境不配合
浏览器实例的视口尺寸是 0x0，所有 getBoundingClientRect() 返回的坐标落在负半轴。点击工具把 rect 中心点当成真实鼠标坐标，直接判定超出可视区域。模型给出的指令是对的，元素也确实存在，失败发生在坐标层。
绕过方式是放弃坐标点击，改为在目标元素上派发完整事件序列：pointerover、pointerdown、mousedown、pointerup、mouseup、click，其中 pointer 前缀用 PointerEvent，其余用 MouseEvent。只发 click 在很多组件库上静默无效，因为内部状态切换挂在 pointerdown 上。
function fire(el, x, y) {
['pointerover', 'pointerdown', 'mousedown', 'pointerup', 'mouseup', 'click'].forEach(function (tp) {
var E = tp.indexOf('pointer') === 0 ? PointerEvent : MouseEvent;
el.dispatchEvent(new E(tp, { bubbles: true, cancelable: true, view: window, clientX: x, clientY: y }));
});
}
工程含义很直接：Agent 的 action 空间不该只有「点击」这一个原子动作，至少要能降级到事件派发，并且能读取环境约束（视口、滚动位置、可见性）。
第二类卡点：框架契约不透明
社区富文本评论框是 contenteditable 的 div。写入要 focus、设置内容、再派发 InputEvent（事件名 input），框架才会把内容同步进自己的响应式模型，字数统计跟着更新。看起来一切正常，但点发送不发请求——提交逻辑挂在键盘事件上，只认 Ctrl+Enter。
var el = document.querySelector('.rich-input');
el.focus();
el.textContent = 内容;
el.dispatchEvent(new InputEvent('input', { bubbles: true, cancelable: true }));
// 提交需要真实键盘事件：Ctrl + Enter
收藏是另一种形式的不透明。点收藏只打开收藏集弹窗，真正的写入发生在第二步：勾选条目内的 checkbox-icon，再点确定，此时才发出 add_article 接口。窄视口下悬浮栏被媒体查询隐藏，还要先强制显形才能取到按钮。
这些行为文档里查不到，只能从 DOM 与网络请求反推。Agent 缺的不是智能，是可观测性：能读网络请求、能读 class 变化，才能把「做了」和「成功了」区分开。
第三类卡点：把隐藏模板读成风控
页面 innerText 里能读到「当前操作失败」，很容易判定为风控而中止。实测那个 risk-modal 常驻在 DOM 里，getBoundingClientRect() 的宽高都是 0，从未弹出——innerText 与 textContent 都会包含隐藏节点。
var m = document.querySelector('.risk-modal');
var r = m.getBoundingClientRect();
// width 为 0 且 height 为 0：模板常驻，并非真实拦截
var real = r.width > 0 || r.height > 0;
判断条件必须是可见性断言：宽度大于 0 才算真弹窗。缺了这一层，Agent 会在完全正常的页面上反复误停；反过来只凭文案就继续，又会把真实拦截吞掉。
第四类卡点：延迟计数与不可控项
社区学习类的浏览任务，单日连开多篇、每篇停留十几秒并滚动，任务页仍可能停在原计数，说明阅读事件上报存在额外限制。反复重试只会制造无效点击，正确做法是记录未计入并在结算里标注。
移动端登录这类项则完全在浏览器 Agent 的能力边界之外，写进「必须手动」清单比绕路实现更省时间。
结论
模型能力决定 Agent 能不能想到下一步；能不能把日常任务跑完，取决于另外三件事：
- action 空间的降级能力：坐标点击之外，有没有事件派发、键盘输入、DOM 直接写入。
- 状态确认机制：每个动作后跟一次断言（字数、class、URL、网络请求），让失败可以稳定复现。
- 可见性与真实性的区分：读 DOM 必须带 rect 判断，否则隐藏模板会持续制造假信号。
按这套流程重跑，成功率提升来自工程改动，而不是换模型。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A counterintuitive observation.
Allowing Agent to operate browsers to perform a set of community routine tasks (browsing, acclaims, collections, comments, publish) is most likely to be attributed to failure: the task was not completed, and the first reaction was often not a model. It is true that every failure is seen in isolation, and the part of the model that is responsible for almost nothing is wrong, but the environment, the framework contract and the confirmation of status.
Category I: Environmental incompatibilities
The view size of the browser example is 0x0, and all the coordinates returned by gettingBendingClintRect() fall on the negative half-axis. Click the tool to use the rect centre as a real mouse coordinate and to determine directly that the visual area is exceeded. The model gives the right command and the elements do exist, and the failure occurs at the coordinates level.
The way around is to drop the coordinates and send a complete sequence of events on the target elements: pointerover, pointerdown, mousedown, pointerup, motionup, click, of which pointer prefix PointerEvent, and the rest with MouseEvent. Only click is not valid on many assembly libraries because the internal state switch is hanging on pointerdown.
function fire(el, x, y) {
['pointerover', 'pointerdown', 'mousedown', 'pointerup', 'mouseup', 'click'].forEach(function (tp) {
var E = tp.indexOf('pointer') === 0 ? PointerEvent : MouseEvent;
el.dispatchEvent(new E(tp, { bubbles: true, cancelable: true, view: window, clientX: x, clientY: y }));
});
}
Project meaning is straightforward: Agent's accelerator space should not be the only atomic action that can be “clicked”, at least down to the event dispatch, and read environmental constraints (vision, rolling position, visibility).
Type II card points: non-transparent framework compact
The community rich text comment box is a capable div. Writing requires Focus, setting content, re-distrieving InputEvent, so that the frame synchronizes content into its own response model, and word count updates. Looks like everything's fine, but no requests are sent at point - submitting logic on keyboard events, only Ctrl+Enter.
var el = document.querySelector('.rich-input');
el.focus();
e. textContent = content;
el.dispatchEvent(new InputEvent('input', { bubbles: true, cancelable: true }));
/ / Submit a true keyboard event: Ctrl + Enter
Collection is another form of opacity. The dot collection opens only the collection window, and the real writing takes place in the second step: checkbox-icon in the tick entry, and then click to make sure that the add article interface is sent. The floating bar below the narrow-view mouth is hidden by media query, and the buttons are not accessible until they are visible.
These actions cannot be found in the files, they can only be reversed from DOM and the network. Agent's lack of intelligence is not an intelligence, it's observability: to read web requests, to read class changes, to distinguish between "do" and "successful".
Number three card points: read hidden templates as wind
The page innerText reads "Current Operation Failed" and can easily be judged to be aborted for wind control. It is measured that the risk-modal is permanently in DOM, that GetBoundingClintRect() is zero in width and never pops up -- innerText and textContent contain hidden nodes.
var m = document.querySelector('.risk-modal');
var r = m.getBoundingClientRect();
/ width is 0 and height is 0: templates are permanent, not real intercept
var real = r.width > 0 || r.height > 0;
The condition for judgement must be a statement of visibility: the width is greater than zero to be true. Without this layer, Agent will stop over and over again on a completely normal page; in turn, it will continue on paper alone and swallow up the real interception.
Category IV card points: delayed count and uncontrolled items
Community-based study-type browsing missions, with multiple single-day sessions, 10 seconds each, and rolling, may still stop at the original count, indicating additional restrictions on reporting reading events. Repeated retrys only create invalid hits, and the correct approach is that records are not recorded and marked in the settlement.
Moving-end login items like this are completely outside the browser's capacity boundaries for Agent, and it is more time-consuming to write "must manual" lists than to bypass them.
Conclusions
Agent's ability to think about next step; it depends on three other things:
- Downgrading capacity of action space: outside coordinates, no event dispatch, keyboard input, DOM writing directly.
- Status confirmation mechanism: each action is followed by an assertion (words, numbers, numbers, URLs, network requests) so that failure can be stabilized.
- Distinction between visibility and authenticity: read DOM must take a rect judgement, otherwise hidden templates continue to create false signals.
The success rate upgrades are based on engineering changes rather than changing models.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.