Manus 引爆通用智能体热潮

2025 年 3 月 6 日，一个叫 Manus 的 AI 产品刷屏了。它的自我介绍很简单：一个通用的 AI Agent，能帮你完成「整件任务」——筛选简历、写分析报告、研究股票、整理资料。演示视频里，Manus 不是逐轮和你对话，而是像真人一样打开浏览器、浏览网页、点按操作、整理结果，最后交付一份完成的成果。

Manus 的发布方式，采用了当时最流行的「邀请码经济」。没有开放注册，想要使用得先拿到邀请码。于是社交平台上一片求码声，二手市场的邀请码被炒到数千元人民币，甚至出现「一码难求」的盛况。服务器在汹涌的流量下过载，官方不得不限流。一个 AI 产品引发如此狂热的抢购潮，此前几乎没有先例。

技术内核上，Manus 属于当时快速成熟的「自主智能体」路线：大模型负责理解和规划，一套工具调用框架让它操作浏览器与各类软件，异步执行让任务可以在后台长时间运行。它不算基础模型上的原创突破，而是把已有的模型、工具、调度组合成了一个能交付完整任务的系统——这种「工程整合」本身，就是 Agent 时代最有价值的能力。

Manus 引发的讨论很快从「好用吗」转向「是真的吗」。有人拿它做测试，有人质疑演示存在选择性，也有人指出它在复杂任务上仍会出错。这些争论并没有掩盖一个事实：它让「把整件任务交给 AI」第一次成为大众可以亲眼看到的现实。2025 年被称为「Agent 元年」，Manus 正是把这个叙事点燃的那根火柴。

热潮过去后，Manus 团队和整个行业都没有停下。中国 AI 创业公司重新聚焦智能体方向，巨头们也把 Agent 列为战略重点。后来的 Manus、OpenAI 的 Agent、各家「通用智能体」产品，都在沿着 2025 年春天这条被 Manus 趟开的路前进。一个产品的热度会退，但它定义的问题——AI 如何接管完整任务——成为整个行业接下来几年的主线。

回看 2025 年 3 月，Manus 的最大贡献，是让「自主智能体」从技术圈的专业术语，变成了大众的集体想象。它未必是技术最强的那一个，但它是最先让所有人看见「AI 可以自己把事做完」的那一个。当人们在 2025 年之后的每一条 Agent 新闻里看到「像 Manus 一样」的描述时，这个春天的这场刷屏，已经悄悄写进了 AI 历史。

On March 6, 2025 a Chinese AI product named Manus flooded the internet. Its self-introduction was simple: a general AI agent that completes "whole tasks" for you—filtering résumés, writing analysis reports, researching stocks, organizing materials. In demo videos Manus did not chat with you turn by turn; it acted like a real person, opening a browser, visiting pages, clicking, organizing results, and delivering a finished product.

Manus adopted the era's favorite launch format: the invite-code economy. There was no open registration; you needed a code to use it. Social platforms filled with people begging for codes, and secondary markets sold them for thousands of yuan, reaching "a code is harder to get than a ticket" proportions. Servers overloaded under the flood, and the company had to throttle. An AI product triggering such a frenzy had almost no precedent.

Technically, Manus belonged to the fast-maturing "autonomous agent" route: a large model handled understanding and planning, a tool-calling framework let it operate browsers and software, and async execution let tasks run long in the background. It was not an original breakthrough in base models but a system combining existing models, tools, and orchestration to deliver complete tasks—and this kind of engineering integration was itself the most valuable capability of the agent era.

The discussion around Manus quickly shifted from "is it useful" to "is it real." People tested it, some questioned the selective demos, others pointed out it still errs on complex tasks. None of that debate obscured one fact: it made "handing a whole task to AI" a reality the public could see with their own eyes. 2025 has been called the "year of agents," and Manus was the match that lit that narrative.

After the hype passed, neither Manus's team nor the industry stopped. Chinese AI startups refocused on agents, and giants made agents a strategic priority. Later general-agent products—from Manus itself to OpenAI's agents—have all followed the road Manus opened in spring 2025. A product's heat fades, but the question it defined—how AI takes over complete tasks—became the industry's mainline for the following years.

Looking back at March 2025, Manus's greatest contribution was turning "autonomous agents" from a technical term into a collective imagination. It may not have been the technically strongest, but it was the first to let everyone see that "AI can finish a task by itself." When every later agent news story describes itself "like Manus," that spring's flood had quietly written itself into AI history.