China's Manus AI triggers viral hype, but real-world tests hit walls

Close your laptop, walk away, and let AI build an entire website or pull market research in the cloud. That was the pitch behind Manus, a new "fully autonomous" AI agent that set the tech world on fire over the past week.

Its official Discord quickly exploded to over 138,000 members, and scarce invite codes were reportedly selling for thousands of dollars on Chinese secondhand app Xianyu. Some influencers were even quick to crown it China’s next "DeepSeek moment." But as early beta testers got their hands on it, reality hit hard.

The hype hits a wall

Manus was launched in preview by a Chinese startup (identified as Butterfly Effect and linked to Monica.im). In a demo video, co-founder and chief scientist Yichao "Peak" Ji claimed Manus is a truly autonomous agent that bridges the gap between idea and execution. He showed it analyzing stocks, ranking resumes into spreadsheets, and claimed it outperformed OpenAI’s Deep Research on the GAIA benchmark.

Its core feature is asynchronous cloud execution. You submit a prompt, shut your device, and specialized sub-agents handle the web browsing, Python scripting, and sandbox execution in the background.

Except when real users tried simple tasks, things went sideways.

A TechCrunch review found Manus crashed after 10 minutes when asked to order a fried chicken sandwich. On a second try, it found the food but failed to complete the checkout. Asking it to book a business flight from New York to Japan only yielded a few broken travel aggregator links. An attempt to build a Naruto fighting game died with an error message after 30 minutes.

Researchers reported similar hiccups. One professor ran Manus alongside OpenAI’s Deep Research on the same prompt. While Deep Research wrapped up in under 15 minutes, Manus choked after 50 minutes at step 18 of 20. Other founders on X noted endless loops and basic factual errors, arguing Manus is basically a prompt-chained workflow rather than a breakthrough agent.

A Manus spokesperson later confirmed that the closed preview is meant to stress-test the system and squash bugs, noting the team is actively adding computing capacity.

Wrapper or real leap?

Unlike DeepSeek, which trained cutting-edge open foundation models from scratch, Manus isn’t building its own foundational layer. Reports indicate it strings together fine-tuned versions of existing models like Anthropic’s Claude and Alibaba’s Qwen via a multi-agent harness.

Hype also spiraled out of control due to viral social media posts. One popular clip showing an agent controlling multiple mobile apps turned out not to be Manus at all, which Ji had to publicly clarify.

Why it matters

In our view, nobody should be paying crazy reseller prices for an invite code right now. As it stands today, Manus looks like an early, fragile prototype that struggles to close the loop on even everyday digital errands.

Yet we think the architecture itself represents the right bet. Moving past the chat box toward asynchronous, cloud-hosted multi-agent systems that run code in sandboxes is exactly where enterprise automation is headed. The massive global interest shows just how eager users are for tools that actually finish work rather than just generating drafts.

The real vulnerability is reliability. When an agent chains dozens of autonomous steps across third-party models, latency explodes and failure rates compound. Until Manus can turn a 50-minute error into a dependable 5-minute execution, OpenAI’s Operator and Anthropic’s Computer Use still hold the upper hand in practical stability.