Thinking about the productization trend of Agent from the prompt content of jailbroken Manus

When AI transforms from a "big talker" to an "execution maniac", a week after Manus became popular, we still need to seriously think about the inspiration this product brings to the development path of the AI ​​industry.

The popularity of Manus is inevitable due to multiple factors. It is like a tireless "digital butler" that can handle cross-domain tasks such as resume screening, stock analysis, travel planning, etc. at the same time, and even complete three days' work while the user is sleeping.

After more than a week, there have been many interpretations or collateral damages at the conceptual level, including saying that Manus is a shell that just uses Claude's advanced API capabilities, or that the invitation code is a marketing hype, etc.

But this does not prevent Manus from having some really cool innovations.

1. Wonderful delivery results. Compared with the text delivery of traditional agents, Manus directly gives super cool web page display;

2. The execution process is extremely transparent. Even if you don’t have an invitation code, you can still experience the entire process of AI execution through the “playback function”;

3. The cost is really low. Compared with OpenAI’s $200/month subscription, Manus only costs $2 to execute a single task.

In fact, the one who was most embarrassed/hurt was OpenAI.

When OpenAI released the o1 slow-thinking model, it did not open its thinking process to the world; instead, DeepSeek’s deep thinking “Well, the user asked me to…”

When OpenAI released the Operator and DeepResearch super agents, they were only available to paying users at $200 per month. As a result, Manus replaced them through the spread of the “replay function”.

Free + good product experience + spreadability can crush everything that seems high-end.

This principle has never changed. There is really a cause and an effect.

There is also a jailbreak article [Manus Tools and Prompts] on Github these days . I suggest you read it carefully, and you can understand a lot of the operating logic behind it. Some key information is as follows:

—— This defines why Manus exists, which is to collect information, create content, and solve problems.

—— This defines the range of Manus’ skills and tools. To put it bluntly, it is what Manus can do.

—— This explains why Manus is a multi-agent system that combines various agents and the tool capabilities mentioned above to thoroughly complete the tasks assigned by users and submit the results.

—— This explains why Manus has a series of special fixed ability modules when performing tasks.

—— This shows that Manus is also capable of using specifications and constraints, including browser rules, coding rules, info rules, and so on.

Therefore, yesterday OpenAI finally couldn't sit still anymore and released the development tool capabilities of the "3+2" suite to developers.

Among them, 3 are three key advanced agent tool capabilities:

• Web Search: Open web search capabilities
  
• File Search: Opening up file and private data search capabilities
• Computer Use Agent: This means that you can browse the web to find information on the cloud side.

The other 2 are advanced integrations of two dialogue feedback modes:

• Response API: From the original chat-only capability to the ability to chat + execute + search

• Agent SDK: From a single agent completing tasks to mobilizing multiple agents to complete tasks collaboratively

This time, OpenAI is not forced into a dilemma like when it was exposed by DeepSeek in February, but it is probably due to some unexpected circumstances.

So they keep emphasizing that 2025 is the year of Agent .

We have to ask, why?

At this point we need to look back at how Manus works.

Manus's transformation goes beyond the traditional AI assistants we have been using, which are apprentices who can only read recipes - if you ask "how to make sweet and sour spare ribs", it can recite the steps, but it cannot turn on the stove or cut meat.

Manus is like a Michelin chef. After receiving the order to "prepare a banquet for eight people", he can automatically buy vegetables, adjust the stove, control the heat, and finally serve delicious dishes.  

This qualitative change stems from three core designs:  

1. Multi-Agent Collaboration

"Conductor" Planning Agent: Like a symphony conductor, it uses the Monte Carlo tree search algorithm (similar to the decision-making method of Go AI) to dynamically disassemble tasks. For example, when processing 100 resumes, it will first scan the file structure, like a chef adjusting the menu according to the inventory of ingredients: prioritize candidates with matching academic qualifications, then analyze work experience, and finally generate a report.  

"Craftsman" Execution Agent: Integrates more than 200 tool interfaces, just like a knife library in the kitchen. It can simulate human operation of browser clicks and scrolls, automatically call Excel to generate charts, and even write data analysis code in Python. Just like a chef controlling a wok, oven, and blender at the same time, the spatula and keyboard become extensions of "fingers".  

"Quality Inspector" Validation Agent: Checks quality through adversarial testing modules. For example, when analyzing financial reports, if the data deviates from the industry benchmark by more than 5%, it will trigger a review process like a food safety detector. This "triple signature" mechanism (three models are independently calculated and then cross-validated) ensures that the results are reliable and avoids serving half-cooked dishes.  

2. "Dynamic recipe" evolution system

Traditional AI "recipes" are fixed, while Manus supports real-time feedback iteration. Users can modify the output format at any time (such as specifying a PPT template), just like when a diner asks for "less salt and more spicy", the chef will immediately adjust the recipe and remember the preference. Even if the task crashes midway (like a power outage while cooking), it can save the progress like a smart refrigerator and use alternative solutions (such as using keyword matching to continue screening resumes) to complete the cooking.  

3. "Aseptic kitchen" safety mechanism

The browser sandbox technology isolates operations, just like processing food in a closed laboratory, to prevent privacy data leakage. The memory preference system is like a personal butler, continuously learning user habits (such as contract template style) and building a personalized strategy library in the cloud.  

Okay, that sounds a bit pretentious.

Simply put, Manus organically combines the excellent paid/free API tools available on the market. The capabilities of each single point are not mine, but the combination of them is the magic of Manus.

While traditional model giants are still creating "models with larger parameters/inference capabilities" for their own use, the Manus team chose to become a "Swiss Army Knife craftsman". By engineering a combination of multiple model tools, the Manus team solved three core problems:  

1. Tool adaptation: Just like allowing a chef to skillfully use kitchen utensils of all brands, Manus develops dedicated drivers for browsers, Office and other software, and can even operate traditional programs without open APIs.  

2. Long-term task management: Traditional AI is like a fast food chef, which must be watched by the user to work. However, Manus supports asynchronous working mechanism, just like a slow cooker can continue to simmer after being removed from the fire, realizing the subversive experience of "user sleeps → AI works overtime".  

3. Cross-border taste fusion: Microsoft Copilot and other products focus on "Western food" (Office scenarios), but Manus can handle "Chinese food" (resume screening) and "French food" (stock analysis) at the same time, breaking through the limitations of vertical fields.  

If Manus can integrate, everyone else can integrate.

When the industry consensus is reached on the supply of the most basic model capabilities, and when the capability boundaries of these models gradually become clear, it is the stage for upper-level applications.

However, can Manus really be called a "general agent" as they define it?

There is a paradox here.

If so, what is the need for other agents? How can we have the year of agent application?

If the same model's API capabilities, the same combination methods and engineering effects are used, what is the difference between Manus1 and Manus2?

The answer is obvious.

My point of view is: models can be universal, but agent products should still flourish.

In different vertical industries, different experience processes are designed for different user groups.

Manus is not an end, but a new beginning.

Despite being close to human performance on the GAIA benchmark (86.5%/70.1%/57.7% on the three-level test), Manus still faces challenges that are common to all kinds of agent products:

1. Limitations of model API capabilities: Some operations rely on the browser environment, just like only being able to cut vegetables with a specific knife  

2. Stability and illusion: Occasional crashes may generate false results. As long as the stability of an API is not 100%, the probability of error will be infinitely magnified when more APIs are combined together.  

3. Uncertainty of business model: On-demand or subscription-based use, these two issues may not lead to a stable user willingness to pay for a long time due to the unresolved issues 1 and 2.

But we still need to fully affirm the value of Manus.

Just as steam engines once leaked air and computers once crashed, every technological breakthrough is accompanied by iteration pains. The value of Manus lies in proving that when AI breaks through the "paper talk" stage and truly becomes an executor who "uses both hands and brain", human-machine collaboration will reach an unprecedented depth.  

In the year of Agent application, more teams can inherit this philosophy of "using both hands and brains". Perhaps we are not far away from the future where "we can change the world just by talking".