Manus: The AI agent that's splitting opinions and setting the industry abuzz

by Wu Yangyu

The AI application boom has finally arrived.

On the evening of March 5, a new AI agent called Manus launched, and within just 24 hours, it became the hottest topic in China’s AI community.

Faced with an unexpected surge in interest, Manus AI’s product lead Zhang Tao, who is also a co-founder of Monica.im, shared his thoughts on social media. "The past few hours have felt like an adventure full of surprises for our team," he wrote.

He admitted that the team underestimated the enthusiasm around Manus. "This was meant to be a phased product update, a moment to share our progress. We only allocated demo-level server resources, never expecting such a massive response."

Still in beta testing and accessible only via invitation codes, demand for Manus quickly outpaced supply. At one point, invitation codes were being resold for tens of thousands of yuan on secondary markets.

A polarizing debate fuels the hype

Opinions on Manus have split into three camps—enthusiastic supporters, skeptical critics, and those advocating for a more measured view.

Media coverage has further fueled the hype, with some dubbing Manus the "DeepSeek moment for AI agents," suggesting it offers a glimpse of Artificial General Intelligence (AGI). Grandiose phrases like "national-level innovation"—once reserved for DeepSeek—have been applied to Manus, drawing even more attention.

But what exactly is Manus? And if it’s not the AI agent equivalent of DeepSeek, what milestone does it represent?

A general AI agent That thinks and acts

Manus bills itself as a "general AI agent," distinct from traditional rule-based, single-task AI assistants. Instead, it integrates multiple AI models and agents to execute complex, multi-step tasks autonomously.

Users can submit a simple command and, after a short wait, receive a polished final product—like a virtual assistant capable of breaking down and executing intricate workflows on its own.

Its product demo showcased several use cases:

• Resume screening: Extracting key details from resumes, ranking candidates, and generating structured reports—including tables when requested.
• Real estate research: Manus evaluates properties by breaking down user preferences—such as safety, schools, and budget—into key factors. It automates filtering through a Python program that cross-references listings with budget constraints, producing a structured report that streamlines decision-making.
• Stock analysis: Fetching financial data, performing correlation analysis, and even deploying a website to visualize findings.

Unlike conventional AI tools that require user oversight at each step, Manus runs asynchronously in the cloud, allowing users to log off while tasks are completed.

The "shell" debate: Real innovation or smart packaging?

A key criticism of Manus is its reliance on external AI models—primarily Anthropic’s Claude—rather than developing its own foundation models. However, the team is upfront about this and doesn’t shy away from the "shell" label.

Tech and AI blogger 01Founder suggests Manus likely incorporates additional reinforcement learning (RL) models—such as Alibaba’s Qwen—and a proprietary tool ecosystem to enhance decision-making. At the heart of its system is a "todo.md" file, which acts as both memory and task manager, ensuring smooth execution and continuity.

Early adopters have tested Manus with creative challenges, from simulating a Google CEO’s journey from startup to tech giant, to organizing messy invoices, and even crafting a minimalist historical timeline of a country, complete with visuals. While results have largely impressed, some limitations remain—such as restrictions on accessing external files and inconsistent aesthetics.

Server stability has also been a challenge, something the team acknowledges as they work to scale up.

Winning with execution, not just technology

Unlike DeepSeek, Manus’s success isn’t built on groundbreaking AI technology but on strong product design and engineering. A Shanghai-based AI industry expert told Jiemian News, "The technology barrier isn’t particularly high—similar ideas began emerging as early as two years ago with the rise of Auto-GPT."

Leading the team are product head Zhang Tao, formerly of ByteDance and Lightyear Beyond, and chief scientist Ji Yichao, founder of Magitech Labs and creator of mobile browser Mammoth. Other key members include Xiao Hong, previously of Wuhan Nightingale Tech, who co-founded the AI productivity tool Monica with Zhang.

Manus’s rise underscores a key industry shift: AI agents are taking off not because of radical breakthroughs, but thanks to better AI model performance and smarter user experience design. Similar concepts have been around since Auto-GPT emerged two years ago, but Manus has succeeded by timing its launch well and making its platform easy to access.

One beta tester summed up the sentiment to Jiemian News: "My experience was smooth overall. There’s nothing particularly new in terms of technology, but the way it’s put together is flawless. This is a true product manager’s moment."

Can AI agents scale, or will they stall?

As AI assistants like Manus gain traction, questions about their business model remain. Some argue that the estimated $2 per task cost is unsustainable. Others believe that with scale, Manus could establish a viable model based on a new metric: Agentic Hours Per User (AHPU).

Investors are watching closely. One AI-focused investor suggests Manus will focus on refining user experience while carefully managing access to prevent server overload and potential reputational damage.

With AI agents poised for a breakout in 2025, Manus will soon face competition from startups like Zhipu AI, which created AutoGLM. Tech giants like ByteDance, Alibaba, and Tencent are unlikely to sit on the sidelines, either.

The Manus team has acknowledged that Manus is still far from the polished experience they aim to deliver in its final version. "There is still significant room for improvement in areas such as model hallucination, user-friendliness of outputs, and processing speed," they stated. They, however, believe they have a three-month head start due to the agility of startups compared to the slower decision-making cycles of large corporations.

As the AI industry braces for its next big leap, the real question is: Who will build the next Manus—and can they scale it successfully?