One command does all the work! Manus AI foreshadows the crazy development of Agents in the next few years

On March 5, an agent system called Manus AI was born.

"The world's first AI agent that delivers complete results!"

In less than 24 hours, the headlines of major technology media have become "Manus AI crushes OpenAI", "Major breakthrough in AI agent technology", "Musk urgently accelerates the development of his own AI agent"...

What exactly is an AI agent? And why is Manus AI so awesome?

Today, Byte Notebook will take you to find out.

Don’t just chat, act independently

Throw away your preconceived notions about ChatGPT — Manus AI is not here to chat with you.

Simply put, you are the boss, and Manus AI is now your subordinate. Give it a task, and it can complete the entire process independently. You no longer need to teach it step by step or whip the AI ​​sentence by sentence.

For example, you can ask for it like this:

"Analyze Tesla stock data for the past 6 months, find price movement patterns, make a beautiful data dashboard, and write an investment recommendation for me."

Then, Manus AI will complete the following steps:

In the above steps, he performed the following tasks in sequence:

• Automatically crawl Tesla stock data
• Writing code to analyze price patterns
• Build a visual dashboard
• Write detailed investment proposal reports

No intervention is required . This is a real AI agent:

Not just to answer questions, but to complete tasks.

According to the official introduction, Manus AI's core capabilities include:

• Autonomous task execution: no need for continuous human guidance
• Multi-tool integration: one-stop service for coding, web browsing, and data analysis
• Real-time monitoring: Check the task progress at any time (worried about it slacking off?)
• Security sandbox: Isolate the code execution environment to prevent "digital runaway"
• Data protection: Encrypted transmission and storage, data is not retained after the task is completed

In the past, you needed a developer, a data analyst, and a content creator to work together for several days to complete a task, but now an AI can do it independently.

This may sound a bit scary, but it is indeed the direction in which AI agent technology is developing.

Is it really better than OpenAI?

Don’t think this is just a marketing gimmick. Manus AI’s performance in the authoritative GAIA benchmark test is indeed amazing!

GAIA is a benchmark specifically designed to evaluate the ability of AI systems to solve complex real-world problems. It contains 466 tasks that require multi-step reasoning.

The average human score on this test is 92%, but how does GPT-4 with plugins perform? A measly 15%.

And the performance of Manus AI? According to reports:

• Level 1 difficulty: 86.5% pass rate (OpenAI's Deep Research only has 74.3%)
• Level 2 difficulty: 70.1% pass rate (OpenAI's Deep Research is 65.8%)
• Level 3 difficulty: 57.7% pass rate (OpenAI's Deep Research is only 47.6%)

Especially on the most difficult Level 3 tasks, Manus AI's performance exceeded OpenAI by a full 10 percentage points .

A gap like this is no longer a gap in magnitude but a generational leap in agency technology.

What can Manus AI do?

What can Manus AI do? According to the internal test preview application, this agent that can work for workers has demonstrated amazing capabilities in the following areas:

I asked Manus AI to take over the refactoring of a Github project. It not only understood the entire code base, but also identified performance bottlenecks, rewrote key components, and even added functional optimizations that I had not expected. The most amazing thing is that it did not need my guidance throughout the process, and the code quality was better than that of the mid-level developers on my team.

I gave Manus AI a bunch of messy Excel files and a vague analysis goal. It automatically cleaned the data, identified key trends, created an interactive dashboard, and gave three business recommendations. This kind of work would have taken me at least 3 days in the past, but Manus AI took only 20 minutes.

Need to prepare teaching materials on the momentum theorem for a high school physics class. Manus AI not only generates lesson plans, but also creates interactive demonstrations, makes quiz questions, and even provides differentiated content for students with different learning styles. It is like a senior teacher with 20 years of teaching experience.

No wonder a venture capital analyst asserted yesterday: "Manus AI is not stealing human jobs, it is creating a whole new category of jobs - 'AI managers'. In the future, we need to learn how to effectively guide AI agents to complete tasks ."

Why can we lead?

According to the research report "Building Effective Agents" released by Anthropic (Claude) last December, successful AI agent systems should have specific design principles. Interestingly, Manus AI seems to fit perfectly:

Simple , Anthropic found that the most successful AI agent implementations do not rely on complex frameworks, but adopt simple, composable patterns. Manus AI is said to use a modular architecture, where each function is optimized independently, rather than a large and comprehensive monolithic system.

A successful agent should clearly demonstrate its planning steps . Manus AI’s real-time monitoring feature allows users to see its “thinking process” and understand why it makes certain decisions.

Tools , Anthropic emphasizes the importance of tool interfaces. Manus AI has put a lot of effort into tool integration. It can not only use tools, but also select the best tool combination according to task requirements .

However, at present, this kind of proxy environment is still isolated in a virtual environment , and it is still some distance away from real productivity. At present, it is only better at collecting and organizing data to call applications or codes. However, I believe that in the future it will soon be directly connected to personal computers to realize a truly automated all-round proxy!