On September 1, 2026, the Chinese startup Manus officially resumed independent operations after the collapse of a deal with Meta worth 2 billion dollars. The company, founded in Singapore, stated that its team will continue to develop general AI agents capable of performing complex multi-step tasks without constant human intervention.
Unlike the previous period of integration with Meta, Manus now emphasizes a shift towards deeper integration into real-world workflows, direct interaction with the outside world, and proactive task execution. This is not merely a return to the status of an independent startup, but a strategic move towards agentic systems that are not limited to text generation but act in real environments through browsers, code, and tools.
In parallel, Recursive Superintelligence, founded by Richard Socher, raised 650 million dollars with the explicit goal of creating recursively self-improving systems. The company aims to automate the entire cycle of scientific research: from identifying model weaknesses to proposing improvements, implementing them, and validating them. Similar rounds were completed by AMI Labs (1,03 billion), World Labs (1 billion), Lila Sciences (550 million), and General Intuition (454 million), indicating a massive influx of capital into agentic AI.
Technically, Manus is an orchestrator that runs frontier models (often Claude) inside an isolated virtual machine with real tools: a browser, file system, and code interpreter. The user sets a high-level goal, and the agent autonomously breaks it down into steps, executes them, checks the result, and returns an artifact. Updates 1.5 and 1.6 showed a reduction in task execution time from 15 to less than 4 minutes and an increase in user satisfaction by more than 19% in blind tests.
The methodology for evaluating such agents remains controversial. The GAIA benchmark, where Manus allegedly outperforms GPT-4, measures reasoning and tool use ability but does not guarantee reliability in long-term scenarios with uncertainty. There is no public data on failure modes when interacting with the real world—for example, errors in web navigation or unstable APIs. This is a typical problem for agentic systems: strong results in controlled tests do not always transfer to production.
Compared to Recursive Superintelligence's approach, Manus emphasizes practical autonomy today, while RSI focuses on a long-term self-improvement loop. Recursive plans a 'Level 1' autonomous system by mid-2026, using millions of dollars on compute. Manus, freed from Meta, can iterate the product faster without corporate constraints but risks falling behind in fundamental recursion research.
DeepMind with AlphaEvolve demonstrates an intermediate path: Gemini proposes solutions, they are automatically tested, the best are selected and improved. This works well in mathematics and code, where results are easy to measure, but is significantly harder in biology or robotics, where experiments are slow and multidimensional. Manus and similar agents try to close this gap through real interactions, but have not yet proven scalability beyond narrow tasks.
Manus's independence allows the company to adapt the product faster to real user scenarios—from competitive analysis to workflow automation. At the same time, the influx of capital into RSI and world models signals a paradigm shift: from scaling language models to creating systems that generate their own experience and improve themselves.
It remains unclear whether agents like Manus or RSI will manage to close the full self-improvement loop without significant human oversight. Independent verification on long-term benchmarks and real-world cases will be the key test for these ambitions in the coming months.
Manus, having returned to independence, gets a chance to accelerate the practical deployment of autonomous agents, but success will depend on the ability to prove reliability in uncontrolled environments.
