- Technical information added
Update March 10, 2025:
X user “Jian” discovered that Manus appears to be using Claude Sonnet with access to 29 tools and the open-source software Browser Use. He found this by requesting the sandbox runtime code from Manus AI. A complete overview of the tools and prompts is available here.
Manus chief researcher Yichao “Peak” Ji confirmed the basic system architecture and provided additional technical implementation details. He explained that the sandbox code is only lightly obfuscated and serves solely to receive commands from the AI agents.
One of Manus AI’s distinguishing features is its multi-agent system, Ji says. Users interact exclusively with an executor agent when communicating with the platform. This agent has no insight into the details of the knowledge agent, planner agent, or other components.
Ad
According to Ji, this architecture helps control context length. It also explains why prompts obtained through jailbreaking typically result in hallucinations, since the executor agent cannot directly access other agents’ knowledge.
While the announcement video (see below) mentioned that Manus AI is powered by “several distinct models,” Ji clarified that it currently uses Claude 3.5 Sonnet v1 and various fine-tuned Qwen models. The team is testing Sonnet 3.7, which Ji says shows promise.
Ji also emphasized that the agent relies heavily on open-source technologies and wouldn’t exist without open source. Manus AI plans to release “quite a few good things” as open source in the near future.
Original article from March 9, 2025:
Manus AI agent could be the next “Deepseek moment” for Western AI labs
A new AI agent called Manus, developed by Chinese startup Monica, demonstrates capabilities in handling complex tasks from travel planning to financial analysis without human intervention. While early demonstrations have attracted attention, the system remains in limited release with key technical details undisclosed.
The Manus website demonstrates how the agent processes real-world tasks, showing examples of its workflow from initial request to final output. The examples include tasks ranging from travel itinerary creation to dashboard building.
The system operates through text prompts – users describe what they want to accomplish, and Manus processes these natural language instructions into completed tasks. For example, when given a prompt about Tesla stock analysis, the system automatically created and published an interactive dashboard to a public URL.
Monica’s co-founder and chief scientist Yichao “Peak” Ji explains in a video presentation how Manus advances beyond traditional chatbots and workflow systems. Unlike tools that only offer suggestions, Manus can execute actions directly through web interfaces. The name derives from “Mens et Manus” (Latin for “mind and hand”), reflecting its dual capability to plan and perform web-based tasks – similar to OpenAI’s Operator and Anthropic’s Claude Computer Use.
According to Monica, Manus outperforms OpenAI’s deep research feature on the GAIA benchmark, which evaluates AI agents on practical tasks. However, the company has revealed little about its underlying technology, making it difficult to verify these performance claims independently.
Initial testing reveals Manus operates in both standard and high-performance modes. This suggests it uses a reasoning model similar to OpenAI’s Operator, which improves its output quality with additional processing time. Operator achieves this through an o3 model that’s been fine-tuned via reinforcement learning specifically for web tasks.
Users have shared examples on X demonstrating capabilities like automated podcast editing. The response has been mostly positive, with some early testers comparing it to Deepseek – another Chinese startup that surprised the industry by matching Western AI capabilities.
Team behind Manus builds on previous AI success
Before developing Manus, founder Xiao Hong established Monica in 2022, creating a browser extension that integrated multiple language models for international markets. The startup secured backing from prominent Chinese investors ZhenFund and Tencent, while co-founder Ji Yichao contributed expertise from his work founding Peak Labs and developing the Magi search engine.
Hong, who goes by “Red,” earned his software engineering degree from Huazhong University of Science and Technology (HUST). His track record includes founding Nightingale Technology, where he developed two AI assistants – “Yi Ban” and “Wei Ban” – that gained over two million corporate users.
The South China Morning Post reports that unexpected demand led to early limitations, raising questions about the system’s capabilities. Product partner Zhang Tao acknowledges their infrastructure was designed only for demonstrations, describing the current version as “still in its infancy, far from what we aim to deliver in our final product.”
Questions remain about the technology powering Manus. While the system likely builds on existing large language models (LLMs), the team hasn’t specified which ones. Similar questions surround Deepseek, another Chinese AI company that reportedly used OpenAI-generated data for training.
The system currently operates as an invite-only web preview, with plans to open-source portions of the technology later this year, acknowledging its roots in open-source development.
The race to develop AI agents
The race to develop autonomous AI agents continues, with mixed progress across the industry. OpenAI has launched its Operator system and a new multi-agent framework called “Swarm”, though early Operator testing revealed significant reliability challenges.
Google’s browser agent Mariner takes a similar approach, incorporating advanced planning abilities and multiple memory types. Industry leaders remain optimistic – both Google DeepMind’s Hassabis and Nvidia’s Huang expect functional agent systems within two years.
Security researchers emphasize caution, however. Recent studies demonstrate that AI agents can be manipulated, creating particular risks when these systems have access to users’ personal web services and accounts.
