Social media has been buzzing with polarizing takes on Manus, an artificial intelligence tool that its creator has described as “a general AI agent that bridges minds and actions.”
Butterfly Effect, the startup behind Manus, clearly holds the tool in high regard. Naturally, industry observers and influencers have taken a keen interest in assessing its potential.
Public opinion has been mixed. Some voices in the AI space have praised Manus as a game changer. Victor Mustar, head of product at Hugging Face, wrote in a post on X that it was “the most impressive AI tool [he has] ever tried.”
Deedy Das, a principal at Menlo Ventures, shared on LinkedIn that Manus helped him analyze Tesla’s stock data and completed two weeks’ worth of “professional-level work” in about an hour.
Amid the debate, KrASIA reviewed online discussions and technical breakdowns to clarify what Manus is—and isn’t—and why it matters.
1. Manus is built for autonomous real-world tasks, not just conversation
Unlike AI models such as OpenAI’s ChatGPT and DeepSeek, Manus is designed to autonomously plan, execute, and complete tasks without human oversight.
The closest comparison would likely be ChatGPT’s Operator feature, which enables similar automation. However, Manus is said to offer broader tool integration and greater autonomy, allowing users to delegate tasks entirely to the AI agent.
2. Its secret sauce? Multi-agent architecture
What sets Manus apart is its sophisticated multi-agent framework. It integrates three specialized agents:
- Planning: This agent leverages Monte Carlo tree search (MCTS) to break down complex tasks and adjust execution strategies in real time. For example, when screening resumes, it might prioritize academic qualifications.
- Execution: Integrated with external tools and workflows, this agent simulates human actions such as clicking, scrolling, and filling out forms. However, it is (currently) limited when working with applications that lack API compatibility.
- Validation: This agent applies adversarial testing to flag inconsistencies. For instance, if financial data deviates from industry benchmarks by more than 5%, it might trigger a review.
Manus also employs multiple AI models in tandem, cross-validating results before selecting the most reliable response based on confidence levels.
3. Manus doesn’t use its own model, yet
Instead of developing its own large language model (LLM), Manus currently relies on external AI models.
On March 10, X user Jian Liao, who is building Pointer, a “general browser agent,” shared that he accessed Manus’ sandbox runtime through prompt engineering. He concluded that Manus leans heavily on Claude Sonnet models and integrates with Browser Use, an AI-powered web automation tool.
Butterfly Effect’s co-founder and chief scientist, Ji Yichao (also known as Peak Ji), later confirmed that Manus utilizes Anthropic’s Claude 3.5 Sonnet and fine-tuned versions of Alibaba’s Qwen. Ji also clarified that while Manus integrates open-source components like Browser Use, it leverages a broader array of external and open-source technologies.
Rather than focusing on model development, Manus appears to prioritize extensive integration, weaving together hundreds of external tools—including browser automation, virtual machines, and dynamic feedback loops.
4. Users have faced early hiccups
While Manus reportedly outperforms OpenAI’s Deep Research on the GAIA benchmark, early users have encountered difficulties.
Alexander Doria, co-founder of AI startup Pleias, posted on X that he ran into error messages and infinite loops while testing Manus. Other users noted that the AI agent struggles with factual accuracy, sometimes failing to cite sources or omitting publicly available information.
A TechCrunch report highlighted additional issues, including Manus crashing while attempting to order a sandwich and providing broken links for flight bookings.
5. There’s risk of replication
By focusing on tool integration rather than foundational AI model development, Manus could be vulnerable to replication by larger competitors such as OpenAI and Google.
Manus seems aware of this risk. It has partnered with Alibaba’s Qwen to accelerate technical improvements and enhance localization for Chinese users.
Ji has also hinted that Butterfly Effect might mitigate competitive pressure by fostering an open ecosystem—encouraging third-party integrations and potentially open-sourcing parts of its tech stack.
Manus remains in an invitation-only phase, accessible via a web interface. It initially launched in English but has since expanded to support nine additional languages, including Chinese, Japanese, and Spanish.
