China’s Manus AI: Autonomous Decision-Making Spark Global AGI Controversy

Manus AI: A Step Toward AGI

On March 5, 2025, Shenzhen-based startup Butterfly Effect, backed by Tencent Holdings and HongShan Capital Group, introduced Manus. The agent operates via a multi-model architecture combining Anthropic’s Claude, Alibaba’s Qwen, and proprietary systems. Its core innovation is CodeAct, a system where actions are executed via Python code generated by LLMs.

Butterfly Effect’s CEO Xiao Hong framed Manus as a “general AI agent” and a step toward AGI, emphasizing its autonomous task execution (e.g., resume screening, real estate analysis): Unlike prompt-dependent AI, Manus reportedly initiates, adjusts, and completes tasks (e.g., resume screening, website creation) with minimal human input, potentially accelerating AGI development.

Practical Applications: Demonstrated use cases include creating playable video games from prompts, designing websites, analyzing stock markets, and screening job resumes.

Performance vs. ChatGPT: In GAIA benchmark testing, Manus provided more detailed responses but took longer to deliver results due to deeper research. Tom’s Guide noted its slower speed but superior depth in head-to-head comparisons.

Early Access Challenges

Limited invite codes (sold on eBay and Goofish/Xianyu) sparked online buzz, but users report crashes, infinite feedback loops, and higher failure rates compared to ChatGPT. Butterfly Effect’s chief scientist, Peak Ji, acknowledged these issues on X.

How Manus AI’s Multi-Agent System Manages Complexity Compared to Single-Agent Systems

Manus AI’s multi-agent architecture represents a paradigm shift in handling complex workflows, addressing limitations inherent to single-agent systems. Here’s a breakdown of its approach and advantages:

1. Task Decomposition and Specialization

• Manus: Uses a central “executor” agent to divide tasks into subtasks (e.g., data retrieval, analysis, formatting) and delegates them to specialized sub-agents (planners, researchers, coders). For example, when creating a real estate report, one agent gathers property listings, another analyzes crime rates, and a third formats insights.
• Single-Agent Systems: Rely on a single model to handle all steps, often struggling with multi-domain expertise. For instance, ChatGPT might generate a basic apartment list but lacks Manus’s ability to cross-reference crime stats and market trends.

2. Parallel Processing and Efficiency

• Manus: Executes subtasks concurrently. When screening job applications, it simultaneously extracts candidate data, assesses qualifications, and compiles spreadsheets—completing tasks in seconds versus hours for humans.
• Single-Agent Systems: Process tasks sequentially, leading to bottlenecks. A monolithic AI might take minutes to generate a financial report, whereas Manus’s parallel workflow reduces latency.

3. Adaptability and Fault Tolerance

• Manus: If a sub-agent fails (e.g., a data fetcher crashes), the system redistributes the task or retries without halting the entire process. This resilience mirrors SmythOS’s analysis of MAS fault tolerance.
• Single-Agent Systems: A single failure (e.g., code execution error) often derails the entire task, requiring manual restart..

4. Scalability for Complex Problems

• Manus: Scales horizontally by adding sub-agents for new domains. During stock analysis, it can integrate Python scripts for data visualization and real-time market APIs without overhauling its core architecture.
• Single-Agent Systems: Require retraining the entire model to add capabilities, limiting agility. For example, expanding Gemini’s skills to website coding would demand significant retraining.

5. Knowledge Sharing and Collaboration

• Manus: Sub-agents share insights via a centralized memory system. When building a physics course, lesson planners and video scriptwriters access shared educational resources, ensuring consistency.
• Single-Agent Systems: Lack mechanisms for internal knowledge transfer, often repeating errors or omitting context.

Comparative Performance

Why This Architecture Matters

Manus’s multi-agent design tackles complexity through collaborative intelligence, mirroring human organizational structures. While single-agent systems excel in narrow tasks (e.g., chess algorithms), they falter in dynamic, multi-domain scenarios. As noted in Forbes, Manus’s ability to autonomously manage workflows like “an efficient employee” positions it as a precursor to enterprise-ready AGI. However, challenges remain in governance and emergent behavior management, as highlighted by Salesforce’s warnings about MAS unpredictability.

Key Innovation: Manus’s asynchronous cloud operation allows persistent task execution (e.g., compiling data overnight), a feature absent in single-agent tools requiring constant user input.

Expert Reactions – Scientific Skepticism

Hugging Face’s Victor Mustar called it “the most impressive AI tool I’ve ever encountered,” suggesting it could redefine coding paradigms.

Critics like Alexander Doria noted factual errors during testing, while AI influencer Langlais criticized its “hunger marketing” tactics and lack of transparency.

A comprehensive survey published in New Scientist on March 14, 2025, reveals deep divisions within the AI research community regarding Manus and similar systems. Of 475 AI researchers surveyed globally, 76% expressed skepticism that current neural network approaches, including those employed by Manus, will lead to genuine AGI.

Main Limitations of Manus AI Identified by MIT Technology Review

1. System Instability and Crashes

• Frequent Crashes: During prolonged usage, Manus experienced system crashes and timeout errors, particularly when handling large text volumes or complex tasks like compiling award nominee lists (MIT Technology Review, March 11).
• Server Overload: Users encountered messages like “Due to the current high service load, tasks cannot be created”, indicating infrastructure strain under demand (MIT Technology Review, March 11).

2. Task Understanding and Execution

• Misinterpretation of Instructions: The AI occasionally misunderstood tasks, made incorrect assumptions, or rushed outputs by “cutting corners” to save time. For example, it initially delivered incomplete journalist lists before refining them with feedback.
• Inconsistent Quality: While adept at analytical tasks (e.g., real estate searches), it struggled with subjective criteria like evaluating “good neighborhoods” (The Download, March 12).

3. Infrastructure and Scalability

• Computational Limitations: MIT’s tests revealed high failure rates compared to tools like ChatGPT DeepResearch, attributed to insufficient server capacity and unstable cloud infrastructure (MIT Technology Review, March 11).
• Cost vs. Reliability Trade-off: Though cost-effective ($2 per task), its infrastructure struggles to support large-scale adoption (36Kr via MIT Review).

4. Restricted Access

• Invite-Only Model: Less than 1% of 186,000+ waitlisted users gained access, limiting real-world testing and feedback.

5. Transparency vs. Reliability

• “Manus’s Computer” Interface: While providing real-time workflow transparency, the interface often froze mid-task, undermining its utility (MIT Technology Review, March 11).

Why These Limitations Matter

MIT’s analysis positions Manus as a promising but immature tool, highlighting gaps in reliability and scalability that must be addressed before it can deliver on its autonomous AI promises. The review concludes it currently functions best as a “highly intelligent intern” requiring close human supervision.

Geopolitical Implications and AI Race

The Manus announcement intensifies the already heated international competition in advanced AI development.

• China’s AI Ambitions: Manus is framed as narrowing the U.S.-China AI gap. Analyst Dean Ball argued it surpasses DeepSeek’s replication strategy, calling it a genuine innovation.
• Privacy Concerns: Users express hesitation about sharing data with a Chinese firm, limiting its adoption for tasks requiring payment or personal details.

Why This Matters

Manus represents a paradigm shift in AI autonomy, potentially accelerating AGI development. Its success could redefine industries by automating complex workflows, but unresolved issues around accuracy, transparency, and trust pose significant hurdles. The polarized reception underscores broader debates about AI’s near-term potential and ethical governance.