Manus AI: The General AI Agent That’s Making Waves

A new AI agent is shaking up the tech world. Manus, developed by the Wuhan-based startup Butterfly Effect, has been making headlines ever since its launch last week. Unlike traditional AI chatbots, Manus operates as a general AI agent, capable of handling a wide range of tasks autonomously.

What Makes Manus Different?

Unlike AI chatbots that rely on a single language model, Manus combines multiple AI models—including Anthropic’s Claude 3.5 Sonnet and fine-tuned versions of Alibaba’s Qwen. This allows it to take on more complex, multi-step tasks. Some have even called it “the second DeepSeek,” comparing it to another unexpected AI breakthrough from China.

Despite all the buzz, access to Manus remains highly limited. Only a small fraction of users on the waitlist have received invite codes. The Manus Discord channel already has over 186,000 members, showing just how much interest this AI agent is generating.

First Impressions: A Smart but Imperfect Assistant

Those lucky enough to get early access describe using  Manus as similar to working with an intelligent and efficient intern. It can sometimes misinterpret requests, make assumptions, or rush tasks, but it responds well to feedback and improves with guidance. English is its default language, and the interface is clean and user-friendly, resembling tools like ChatGPT and DeepSeek.

One of  Manus’s standout features is its “ Manus’s Computer” window, which lets users watch the AI in action and even step in if needed. This transparency makes it feel more collaborative than other AI tools.

Putting Manus to the Test

To evaluate Manus’s capabilities, it was given three tasks:

1. Finding Journalists Covering China Tech

Initially, Manus provided a small list of names, with some missing key details. When asked why, the AI admitted it had rushed the process to save time! After further instructions, it produced a much more comprehensive list of 30 journalists, complete with their affiliations and notable work. It also made it easy to download the results as a Word or Excel file.

A key limitation, however, was its struggle with paywalled content. Since many media sites block automated tools, users had to manually step in to complete some research tasks.

2. Searching for Two-Bedroom Apartments in NYC

 Manus was asked to find listings that met specific criteria, including budget, location, and amenities. Initially, it interpreted “outdoor space” too strictly, only showing properties with terraces or balconies. But after some clarification, it refined its search and presented results in a well-organized format, complete with categories like “best overall,” “best value,” and “luxury option.”

This task took under 30 minutes significantly faster than the journalist search—since real estate listings are more structured and widely available online.

3. Selecting Innovators Under 35

The most challenging test was asking Manus to compile a list of 50 candidates for MIT Technology Review’s Innovators Under 35 list. Manus started by studying past selections, then developed a search strategy to find potential candidates. However, it ran into issues accessing academic papers and paywalled content.

After three hours, it had only found three fully researched candidates. When pushed to provide a complete list of 50, it did—but with a clear bias toward well-known institutions. When asked to include five candidates from China, it managed to do so but leaned heavily on high-profile media figures. Eventually, the system reached its limits and began to struggle with performance.

Strengths and Weaknesses

 Manus shows a lot of promise, but it’s not without its flaws. On two of the three tasks, it outperformed ChatGPT DeepResearch in quality but took longer to complete them. It excels at structured, research-heavy tasks but can struggle when working with paywalls, complex academic searches, or very large data sets.

Some users have reported system crashes, slow response times, and temporary task limits due to high demand. However, Manus’s per-task cost is reportedly only $2 one-tenth of the cost of DeepResearch. If its developers can improve stability, it could become a go-to tool for professionals, independent developers, and small teams.

The Future of AI Agents

One of  Manus’s best features is its transparency. Users can see how it works, intervene when needed, and even replay past sessions. It remembers important instructions, making it highly customizable.

While Manus still has room for improvement, it represents a significant step forward in AI agents. It’s more than just another chatbot it’s an AI assistant that can think and work like an intern. As AI development continues, we can expect more tools like Manus to reshape how we approach research, automation, and complex decision-making.