Manus: Three Days of Real Usage — Here's My Honest Take

Remember when Manus first went viral?

Invitation codes selling for 100,000 yuan. Social media flooded with “code begging” posts. Tech media breathlessly reporting that “a Chinese team created the world’s first general AI agent.” I wrote an analysis at the time, but honestly — no code, no hands-on experience, could only evaluate based on others’ review videos.

Now it’s different. Manus opened registration to everyone. I signed up immediately and used it for three straight days. Today I’m sharing my honest thoughts.

First, the conclusion: It works, but not as magically as imagined.

Where does it work well?

I tested about 20 tasks. Manus performs well in these scenarios:

First is information organization. For example: “Help me organize important AI industry news from the past week, categorized by company, no more than 3 items per company.” Manus can open web pages, search, filter information, and organize it into tables. This process is truly autonomous — no step-by-step guidance needed.

Second is simple data processing. I threw it a CSV file and asked it to “analyze sales trends and find the three fastest-growing categories.” It can write Python code to run the analysis and give conclusions. Code quality is average, but usable.

Third is workflow tasks. For example: “Help me find 5 open-source RAG projects on GitHub with over 1000 stars, sorted by update time.” Tasks with clear goals and well-defined steps, Manus completes quite well.

Where does it fall short?

But honestly, there are more failure cases.

The most typical is tasks requiring deep business logic understanding. I asked Manus to “analyze our customer service conversation records from last month and identify the main pain points of user complaints.” It could open the file and run some keyword statistics, but the conclusions were superficial — “users mentioned ‘slow’ frequently” — as if that wasn’t obvious.

The problem is, it can’t understand what “slow” means in different contexts. Is it slow page loading? Slow customer service response? Or slow logistics? Manus can’t distinguish these.

Another issue is long tasks tend to break. I asked it to “help me research competitor Company A’s product features and compile a comparison report.” After running for 15 minutes, it just stopped. No error message, just stopped. I suspect token limits or timeouts, but users have no idea what happened.

Gap with early demos

Remember the demo videos Manus released early on? That demonstration of “automatically screening resumes, scheduling interviews, and sending emails” was indeed impressive at the time.

But after actual use, I found those demos had one thing in common: clear task boundaries, definite input and output.

Real-world work is rarely like this. Most tasks are vague, require repeated communication, and requirements may change midway. Manus’s performance is less stable when facing this “real-world complexity.”

This isn’t Manus’s problem — it’s a common problem for all “general AI agents.”

Is the general agent a pseudo-proposition?

After three days using Manus, I’ve been thinking: Is the concept of “general AI agent” itself overhyped?

My view: Not a pseudo-proposition, but don’t expect it to replace humans in complex work anytime soon.

Current Manus is more like an “enhanced automation tool” — more flexible than traditional RPA, can handle some vague instructions, but far from “understanding and executing arbitrary tasks like a human.”

Truly usable scenarios are still those with clear goals, decomposable steps, and high fault tolerance. Like information collection, simple data processing, and workflow operations.

Is it worth using?

If you were curious about Manus before but couldn’t get a code, you can try it now. The free quota is enough to play with for a while, at least enough to form your own judgment.

But don’t set expectations too high. It’s not an “AI employee,” more like “a somewhat smart but not very reliable intern” — can help with some chores, but important matters still need to be handled yourself.

My personal feeling: Manus represents a direction for AI agents, but the endpoint of this direction may not be “one general agent that can do everything,” but rather “many specialized agents for specific tasks, working together in some way.”

What do you think? Have you tried Manus? How was your experience?