Don't Teach AI, Just Throw Tasks! The Manus Team Tears Off the Emperor's New Clothes of AI Agents

In March 2025, an AI Agent product named Manus suddenly swept through social media. Without any preheating, marketing, or even a stable server (which was overwhelmed to the point of requiring invitation codes), everyone was discussing it. From tech bloggers to ordinary users, from investors to competitors, the topic was the same: "What exactly did Manus do?"

The answer might lie in its slogan: "Manus has no secrets."

From L1 to L3: The "Seeing" Revolution of AI Agents

Manus' core narrative is simple: The evolution of AI Agents is a leap from L1 (basic interaction) to L3 (complex autonomous tasks). However, its disruptiveness lies in "seeing".

• L1 Era: Users had to learn "prompt engineering," taming AI like a trainer with fixed commands.
• L2 Era: Chain-of-Thought (COT) emerged, allowing AI to break down tasks, but still requiring humans to design the workflow.
• L3 Era: Manus' answer is "Less structure"¡ªdiscarding preset frameworks and letting AI autonomously plan, experiment, and deliver like a human intern.

A typical case: Manus was tasked with "counting the types of penguins appearing simultaneously in a YouTube video." Instead of relying on pre-programmed visual models, it directly operated video shortcuts (press K to pause, press 3 to skip), took dynamic screenshots, and provided the answer. This ability to "see the task and solve it autonomously" allows non-technical users to intuitively feel the "intelligence" of AI.

Technology "No Secrets": The Brutal Beauty of Less Structure

The Manus team openly stated: "We didn't write any preset workflows." This is backed by two counter-intuitive principles:

1. No Teaching, Just Incentives:
   Adopting the reinforcement learning philosophy of AlphaZero, Manus' model explores solutions through "trial-and-error rewards" rather than relying on human experience. As the technical lead said, "Adding more rules only limits the model, like stuffing an intern with an operations manual, which stifles creativity."
2. Generalization Over Specialization:
   Manus rejects shortcuts in vertical scenarios and insists on using the same model to handle tasks across all fields, from data analysis to game development. The cost of this "brutal generalization" is initial instability, but the potential is higher¡ªjust like in the AI 2.0 era, a generalist eventually outperforms countless specialists.

Product Magic: Letting Users "See" AI's Thinking

Manus' killer feature is the transparency of AI's workflow. For example, when processing the task of "analyzing Chinese companies on the a16z list," users can see the agent's complete thought process:

• First, search the web for candidate lists
• Investigate each company's background one by one
• Cross-verify ambiguous information
• Finally, compile the report

This "intern-like" real-time progress display eliminates the "black box" feeling of AI and even gives users the illusion of "collaborative participation." As early testers reported, "It doesn't feel like a tool, but more like a clumsy yet reliable colleague."

Controversy and Bold Claims: Shell Games, Open Source, and "Pseudo-Innovation"

Manus' popularity also attracted skepticism:

• "Isn't this just a DeepSeek shell?"
  The team responded: If connecting to the same large model counts as a shell, then Perplexity and Cursor should also be categorized as such. The real barrier lies in how to release the model's potential through product design.
• "Open-source and replicate in 3 hours!"
  Indeed, some teams quickly launched "OpenManus," but users soon found out: Deployment required command lines, API calls, and server setups¡ªwhile Manus' strength was its "plug-and-play" user-friendly experience.
• "Big Tech will crush it!"
  However, Manus' CEO, Zhang Tao, was optimistically contrarian: "If the giants don't follow, it means we've chosen the wrong track."

The "Accidental" and Inevitable Marketing

Manus' spread is a textbook case of a "zero-budget hit":

• Seed User Leverage: Early invitation codes were deliberately limited, which paradoxically triggered a "scarcity frenzy" on social media.
• Anti-Routine Narrative: The technical team's closed-door sharing PPT (the document analyzed in this article) was leaked, and the hard-core content unexpectedly went viral.
• Emotional Resonance: In 2025, amid widespread AI anxiety, a "visible" product naturally became a hot topic.

As one investor lamented, "You can't replicate this kind of spread, just like you can't replicate a forest fire."

Manus may fail, but it has revealed an irreversible future for AI Agents:

When technology has no secrets, the key to victory becomes¡ª who can package "general intelligence" into a "wow moment" at users' fingertips.

In this revolution, the biggest winners might be those "who can't code but are good at asking questions." Because the future competition is no longer about "who owns AI," but "who can be AI's boss."