Inside Manus: China’s New AI Agent That’s Turning Heads

The buzz around the new general AI agent Manus has been loud and global. Developed by Wuhan-based startup Butterfly Effect, Manus has drawn major attention from the tech world, including shoutouts from Twitter co-founder Jack Dorsey and Hugging Face’s Victor Mustar. It’s even earned comparisons to DeepSeek, the previous surprise hit in AI. But does it live up to the hype?

Despite the excitement, few users have gotten past the waitlist—less than 1%, in fact. Still, the interest is real: the official Discord has over 186,000 members, and Manus is being hailed as one of the first AI tools to act like a true general-purpose agent, not just a chatbot. Unlike single-model systems like ChatGPT or DeepSeek, Manus operates by coordinating multiple AI models (including Claude 3.5 Sonnet and custom versions of Qwen) and autonomous agents to perform tasks independently.

Testing the AI Intern

MIT Technology Review got early access to Manus and tested its capabilities. Using it felt surprisingly natural, like working with an intelligent, responsive intern. The tool sometimes misunderstood instructions or tried to speed through tasks, but when prompted with detailed feedback, it quickly adjusted. Its design is clean, user-friendly, and optimized for English-speaking users.

Upon logging in with an invite code, users land on an interface similar to ChatGPT’s, chat input in the center, session history on the left, and curated task suggestions on the main page, including business planning, educational tools, and personalized wellness routines.

A standout feature is the “Manus’s Computer” window, which shows you everything the agent is doing—step by step. You can watch, pause, or jump in to assist. This transparency is rare among agentic AI tools and makes the experience feel much more collaborative.

Task Performance: From Journalism Lists to Real Estate

Three main tasks were used to evaluate Manus:

1. Finding Top China Tech Journalists
   The initial list was short and inconsistent. When asked why, Manus admitted it had rushed. After requesting more depth and clarity, the agent returned with a much richer, well-structured list of 30 journalists, including outlets and key articles. It even allowed downloads in Word or Excel formats. The downside? It struggled with paywalled content and captchas, needing human help in those cases.
   
2. Searching for NYC Apartments
   Manus was given detailed criteria—budget, outdoor space, location, etc. At first, it interpreted “outdoor space” too narrowly, excluding many good options. But after refining the request, it generated a useful tiered list, complete with labels like “Best Value” and “Luxury Pick.” The process took under 30 minutes—much faster than the journalism task.
   
3. Nominating Innovators Under 35
   This was the most complex assignment. Manus began by studying past nominee profiles and built a search plan. After three hours, it produced only a few full profiles and a rough list of 50 names, many skewed to a few regions and sectors. When asked for five Chinese candidates, it delivered, but leaned toward well-known media figures. The system also warned that processing too many prompts could degrade performance.

Not Perfect, but Promising

Overall, Manus performs best on focused research tasks with open-source data. Its intern-like nature makes it especially useful for professionals needing structured outputs, like lists, summaries, or research briefs. While it outperforms ChatGPT’s DeepResearch in quality, it can be slower and less stable, often crashing or freezing during tasks. A high server load occasionally prevents new sessions from starting.

But Manus’s price point is a major plus: $2 per task, compared to $20 for DeepResearch. And according to its creators, server improvements are on the way.

Final Verdict

Manus’s real strength lies in its transparency, adaptability, and collaborative design. Users can follow along, step in when needed, and even influence the agent’s learning over time. Each session is saved and replayable, making it easy to share or review progress.
While it still needs fine-tuning, especially with large tasks and stability, Manus signals that China’s AI ecosystem is not just catching up, but innovating in its direction, pushing the boundaries of what autonomous AI can do.