OpenAI 发布 Operator

2025 年 1 月，OpenAI 发布了一个叫 Operator 的产品。它的能力听起来像科幻片：一个 AI，能自己打开浏览器，登录网站，填写表单，完成预订——从订餐、订票到购物，全程不需要你动手。这不再是你问一句它答一句的聊天，而是「你下达任务，它替你跑腿」的通用 Agent。

Operator 背后的技术思路叫 Computer-Using Agent，核心是让模型像人一样「看屏幕、动鼠标、敲键盘」。它运行在云端浏览器里，用户的每一次操作都能被观察：AI 在做什么、走到哪一步，你随时可以喊停、接管。面对支付、登录这类敏感操作，它设置了确认机制，必须经过用户同意。这种「可观察、可接管、可确认」的设计，是早期 Agent 产品最在意的安全底线。

发布之初，Operator 最被关注的问题有两个：一是它到底能不能可靠地完成真实任务——毕竟网页千奇百怪，登录墙、验证码、动态布局都是考验；二是「让 AI 拿着你的账号操作」这件事，风险边界在哪里。Operator 的答案是用「监督式自主」来平衡：能自主，但不盲目自主。

Operator 的意义，在于它把「Agent」从一个演示概念变成了一款可以订阅使用的产品。2025 年被称为「Agent 元年」，而 Operator 是第一个把「AI 替我在网页上办事」做成大众化服务的产品。它之后，Manus、各家浏览器 Agent 层出不穷，整个行业的叙事从「会聊天的模型」转向「会干活的模型」。

回看 Operator 的发布，它开启的是一场交互方式的革命：从「人和软件交互」变成「人委托 AI 和软件交互」。今天各种 Agent 产品已经层出不穷，但回看起点，是 2025 年 1 月这个自己会操作浏览器的产品，第一次让「AI 替你干活」变成了可以买到手的现实。

In January 2025 OpenAI released a product called Operator. Its capability sounded like science fiction: an AI that opens a browser by itself, logs into sites, fills forms, and completes bookings—ordering food, tickets, shopping—without you touching anything. This was no longer chat where you ask and it answers; this was a general-purpose agent where you delegate a task and it runs errands for you.

The idea behind Operator is called a Computer-Using Agent: making the model behave like a person—look at the screen, move the mouse, type. It runs in a cloud browser, and every step is observable: what the AI is doing, where it is in the task, and you can stop or take over at any time. For sensitive operations like payments and logins, it has confirmation flows requiring your consent. That design of "observable, interruptible, confirmable" was the safety floor early agent products cared about most.

Two questions dominated the launch. First, could it reliably complete real tasks—the web is chaotic, with login walls, CAPTCHAs, and shifting layouts. Second, where is the risk boundary of letting an AI operate with your accounts? Operator's answer was "supervised autonomy": autonomous, but not blindly so.

Operator's significance was turning "agent" from a demo concept into a subscription product. 2025 was called the year of the agent, and Operator was the first mainstream service where "AI does my web tasks" became buyable reality. Manus and countless browser agents followed, and the industry narrative shifted from "models that chat" to "models that do".

Looking back at Operator's release, it opened a revolution in interaction: from "people interact with software" to "people delegate AI to interact with software". Today agents are everywhere, but at the origin is this product from January 2025 that could operate a browser itself—the first time "AI does it for you" became purchasable reality.