Just before sunrise on March 6, 2025, a select group of beta testers watched an AI agent called Manus accomplish feats that felt like science fiction. In quick succession, it unzipped résumés, scanned their contents, wrote Python scripts to analyze stock data, and even launched a rudimentary website—without pausing to ask for help. For a moment, it seemed like the future had arrived early.
Manus bills itself as a “general-purpose AI agent”—one that not only processes information but also acts on it. Users provide a goal, and Manus handles the intermediate steps on its own, sifting through data, writing code, and making decisions. According to early demos, it can build travel itineraries and book flights, screen job applications, and conduct financial modeling while you focus on something else entirely.
Its creators, a Beijing-based startup known informally as Monica.im, say the name “Manus” reflects their aim to merge “mind and hand.” In other words, they envision an AI that moves beyond thinking to doing—a digital colleague that proactively tackles multi-step tasks without constant nudging or oversight.
The development team claims Manus has broken records on the GAIA benchmark, a test that evaluates how well AI agents solve complex, real-world challenges. Preliminary scores appear to outpace previous high watermarks set by top-tier models, including some from OpenAI. However, these findings are based on internal testing, with little external review. Observers note that independent validation remains crucial, especially given how competitive the AI landscape has become.
Despite the ongoing need for corroboration, excitement soared when Monica.im offered limited beta invites. Some codes reportedly resold for thousands of dollars on secondary markets, and social media lit up with both eager praise and wary skepticism. Company representatives insist they underestimated user demand—and their servers’ capacity.
Manus’s roots trace back to its lead developers, Xiao Hong and Ji “Peak” Yichao. Both are thirty-something tech entrepreneurs with track records in software innovation. Xiao Hong founded successful productivity tools that earned backing from industry heavyweights, while Ji gained recognition for a novel web browser. Supported by investments from Tencent and ZhenFund, their company, Butterfly Effect, created Monica.im as a multi-model AI assistant. Manus is the logical next step in their quest: building an agent that not only informs but also executes.
As buzz around Manus grew, the team emphasized that the public demonstration represented a beta-stage product. They freely admit to capacity issues and occasional slowdowns. Beta testers describe a range of experiences: some tout smooth performance, while others claim tasks bog down under heavy loads. Such reports confirm that while Manus’s potential may be vast, it still needs room to mature.
New capabilities often spark fresh concerns. Because Manus can operate asynchronously in the cloud, some worry about the ethical and security implications of a largely unattended AI. Could malicious actors co-opt it to launch cyberattacks or manipulate data? How do developers guard against unintended consequences if Manus is entrusted with sensitive tasks in finance, healthcare, or defense?
Open-sourcing parts of Manus’s code—slated for later this year—may improve transparency by allowing external experts to spot vulnerabilities or bias. Yet it also raises the question of whether adversaries could repurpose the technology. Balancing open collaboration with the need for security remains a delicate challenge.
Whether Manus fulfills its transformative promise depends on rigorous testing, thoughtful safeguards, and continued transparency. If independent benchmarks corroborate the results so far, we could see a new era of AI assistants that free people to focus on strategy and creativity rather than routine drudgery. Already, some hail Manus as the leading symbol of China’s drive to compete globally in advanced AI research.
Others caution that the real measure of Manus’s impact will come once more users wield it under real-world conditions. As testing expands and scrutiny increases, either we’ll see Manus’s capabilities validated—or discover painful shortcomings. In the meantime, the project’s rise highlights a broader truth: the gap between AI “thinking” and “doing” is closing. Whether that spells an era of unparalleled productivity, job displacement, or a mix of both remains a subject of lively debate.