Manus AI: Claude wrapper with browser tools hyped as 'the world's first fully autonomous AI agent'

Summary

Manus, a Chinese AI agent company, launched March 6 claiming to be the world's first fully autonomous AI agent. A Forbes contributor profile compared it to DeepSeek's impact and suggested China had leapfrogged Western capabilities in autonomous systems.

The product is built entirely on Anthropic's Claude Sonnet model. It wraps Claude with 29 tools and integrates Browser Use, a YC-backed browser control product. When developers asked Manus for its own runtime code, the system revealed it was Claude Sonnet with tooling on top. The company appears to have intentionally obscured that Browser Use was the underlying browser-control layer.

The Forbes article reads like unvetted PR. It opens with the standard Silicon Valley profile setup before claiming Manus doesn't require human prompts and can execute end-to-end workflows independently. The actual value proposition is more concrete. Manus can take a zip file of resumes, extract skills, cross-reference job market trends, and present hiring recommendations. It can scrape social media, generate a biography, code a website, and deploy it without follow-up prompts. These are useful capabilities, but they depend on existing models gaining access to real tools, not on breakthrough architecture or model capabilities.

Philip Schmid, a researcher at Google DeepMind, observed that Manus outperforms OpenAI's Deep Research on one benchmark. The core insight is that building AI products does not require training a proprietary foundation model. Vast low-hanging fruit remains in giving existing models access to Python, web browsers, and databases. Models like Claude Sonnet stay static in their core capability and perform better with better tools and integrations.

The framing of Manus as a Chinese breakthrough misses the facts. It relies on Anthropic's model (a UK-founded company with a San Francisco presence) and open-source browser tooling. The product itself, with multi-agent architecture, asynchronous background task execution, and isolated sandbox environments, represents solid engineering but not a paradigm shift. The timeline narrative that China is leading autonomous systems does not hold up.

What Manus actually signals is that competition is moving from model weights to product integration. Companies that can integrate existing models into real-world tools like browsers, code execution, and databases will outperform those shipping raw model interfaces. That pattern is real. But it does not require a Chinese foundation model or represent a geopolitical turning point. The article's framing obscures a simpler story: competent product teams can wrap commodity models into useful agents. That is happening everywhere.