Manus has quickly become one of the more closely watched AI agents because it claims to move beyond chat and into direct task execution. The system, developed by Chinese startup Monica, has been shown handling requests such as travel planning, financial analysis, dashboard building, and automated podcast editing through natural language prompts.
New technical details add a clearer picture of what is behind those demonstrations. According to information surfaced from Manus and later confirmed in part by Manus chief researcher Yichao "Peak" Ji, the agent relies on Claude 3.5 Sonnet v1, various fine-tuned Qwen models, open-source technology, and a multi-agent architecture designed to divide work across separate components.
The Technical Picture Is Becoming Clearer
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 was also made available.
Ji confirmed the basic system architecture and added more implementation context. He said the sandbox code is only lightly obfuscated and exists solely to receive commands from the AI agents. That matters because it suggests the visible runtime is not the whole intelligence layer, but rather a command-receiving environment connected to a broader agent system.
The earlier announcement video said Manus AI is powered by "several distinct models." Ji later clarified that the current setup uses Claude 3.5 Sonnet v1 along with various fine-tuned Qwen models. The team is also testing Sonnet 3.7, which Ji says shows promise.
Open source is central to the story. Ji emphasized that Manus AI depends heavily on open-source technologies and said it would not exist without open source. Manus AI also plans to release "quite a few good things" as open source in the near future.
Why Manus Is Not Just Another Chatbot
Manus is presented as an agent that can plan and act through web interfaces rather than simply answer questions. Users describe a goal in text, and the system attempts to turn that request into a completed task. The Manus website shows examples of workflows from initial request to final output, including travel itinerary creation and dashboard building.
One example involved a prompt about Tesla stock analysis. In that case, the system automatically created and published an interactive dashboard to a public URL. Other examples shared by users on X include automated podcast editing.
Monica's co-founder and chief scientist Yichao "Peak" Ji has described Manus as a step beyond traditional chatbots and workflow systems because it can execute actions directly through web interfaces. The name comes from "Mens et Manus" (Latin for "mind and hand"), reflecting the combination of planning and web-based action.
The comparison point is clear: Manus is being discussed alongside OpenAI's Operator and Anthropic's Claude Computer Use. The shared theme is that AI systems are moving from advice into execution, especially inside browsers and web services.
The Multi-Agent Design Explains Some Limits
One of the most important details is Manus AI's multi-agent system. Ji says users communicate only with an executor agent. That executor agent does not have visibility into the knowledge agent, planner agent, or other internal components.
This design has a practical benefit: it helps control context length. Instead of putting everything into one agent's view, the system splits responsibilities. But it also explains why prompts obtained through jailbreaking often produce hallucinations. The executor agent cannot directly access the knowledge held by other agents, so attempts to force it to reveal hidden system details may produce unreliable answers.
That architecture also makes outside evaluation harder. If users only see one part of the system, they cannot easily inspect how the other parts reason, plan, or retrieve information. For an AI agent that claims to perform complex work without human intervention, that opacity is a major part of the evaluation challenge.
Performance Claims Still Need Caution
According to Monica, Manus outperforms OpenAI's deep research feature on the GAIA benchmark, which evaluates AI agents on practical tasks. But the company has revealed limited information about the underlying technology, making independent verification difficult.
Initial testing showed that Manus operates in both standard and high-performance modes. That suggests it may use a reasoning model similar to OpenAI's Operator, which improves output quality with additional processing time. Operator achieves this through an o3 model that has been fine-tuned via reinforcement learning specifically for web tasks.
Early reaction has been mostly positive, and some testers have compared Manus to Deepseek, another Chinese startup that surprised the industry by matching Western AI capabilities. The phrase "Deepseek moment" has become part of the discussion around whether Manus could pressure Western AI labs in a similar way.
Still, the product remains in limited release. Manus currently operates as an invite-only web preview, and key details are still not fully disclosed. The South China Morning Post reports that unexpected demand created early limitations. Product partner Zhang Tao described the current version as "still in its infancy, far from what we aim to deliver in our final product."
The Team Behind Manus
Before Manus, founder Xiao Hong established Monica in 2022. The company first created a browser extension that integrated multiple language models for international markets. Monica secured backing from prominent Chinese investors ZhenFund and Tencent.
Hong, who goes by "Red," earned a software engineering degree from Huazhong University of Science and Technology (HUST). He previously founded Nightingale Technology, where he developed two AI assistants, "Yi Ban" and "Wei Ban," that gained over two million corporate users.
Ji Yichao brought experience from founding Peak Labs and developing the Magi search engine. That background helps explain why Manus is being framed not merely as a chatbot, but as a system for completing practical web-based tasks.
The Bigger Agent Race
Manus enters a wider race to build autonomous AI agents. OpenAI has launched Operator and a new multi-agent framework called "Swarm," though early Operator testing revealed significant reliability challenges. Google's browser agent Mariner follows a related path, using advanced planning abilities and multiple memory types.
Industry leaders remain optimistic. Google DeepMind's Hassabis and Nvidia's Huang expect functional agent systems within two years. But security researchers urge caution, since studies show that AI agents can be manipulated. The risks become sharper when agents have access to personal web services and accounts.
For now, Manus is both a signal and an unresolved question. Its use of Claude 3.5 Sonnet v1, Qwen models, Browser Use, and 29 tools makes it less mysterious than before. But until broader access and more technical disclosure arrive, its most important claims will remain difficult to judge from the outside.