86% Human: The breakthrough architecture closing the AI-human gap

Something revolutionary is introduced in the AI world, and a new competitor is emerging, generating significant buzz across the tech industry.

Manus AI, possibly the most sophisticated AI agent available today, was developed by Chinese startup Monica and launched on March 6, 2025. The buzz in the AI space is that it is like "China's next Deepseek moment."

Manus positions itself as an actual general-purpose AI agent capable of autonomously handling various complex tasks, unlike regular chatbots like Chatgpt, Claude, etc.

It can autonomously handle tasks from travel planning and financial analysis to searching through dozens of files and conducting industry research.

Manus doesn't rely on a single large neural network. Instead, it has its innovative multi-agent architecture. Manus functions more like a manager who divides a task into sub-tasks and distributes it to specialised sub-agents.

"From day one, we decided to work orthogonally to model development, wanting to be excited rather than threatened by each new model release," explained Yichchow Peak G, co-founder of Manus AI.

The system has 29 different integrated tools, such as specialised sub-agents, to automate web navigation, run code securely, and extract critical information from files.

Anthropic's Claude 3.7 Sonnet model powers Manus as the central system, complemented by open-source technologies like YC company's browser tool and E2B's secure cloud sandbox environment.

Manus has scored 86.5% on the GAIA(General AI Assistants) benchmark, which tests AI agents on reasoning, multimodal handling, web browsing, and tool proficiency.

A normal human being would score 92% on this benchmark, and the fun fact is that it significantly outperforms competitors like OpenAI's Deep Research, which scored around 74%.

The potential applications of Manus AI are huge. Early users report success with creating detailed travel itineraries, conducting financial analyses, developing educational content, compiling structured databases, comparing insurance policies, sourcing suppliers, and even assisting with high-quality presentations.

However, there are also controversies about the system; some critics have dismissed Manus as merely an "AI wrapper"- a service that utilises current foundation models with various tool calls rather than building something revolutionary.

This criticism overlooks the fact that many successful AI products today, including code assistants like Cursor and specialised legal tools like Harvey, follow a similar approach.

The distinctive characteristics of AI platforms aren't necessarily creating everything from scratch but how effectively they combine existing technologies with intuitive user interfaces, proprietary evaluations, careful fine-tuning, and thoughtfully designed architectures.

Manus has several advantages in this regard. Its multi-agent orchestration costs only $2, which is significantly lower per task cost compared to integrated competitors.It also offers greater transparency and user control, allowing users to inspect, customise, or replace individual sub-agents and tool integrations.

Despite these strengths, Manus has its limitations. As the task becomes more complex, coordination across specialised agents becomes increasingly tricky. Additionally, competitors could commoditise its current advantages in user experience, targeted fine-tuning, and thoughtful integrations.

Currently, it is in the beta testing stage, and it has already faced some early reports of glitches and performance inconsistencies. However, these are common problems for new, growing revolutionary technologies.

As AI evolves from reactive systems to autonomous agents capable of making independent decisions, platforms like Manus represent a significant step forward.

We still don't know if it will truly become the definitive breakthrough in AI agents. Still, its innovative approach to complex problem-solving through coordinated multi-agent systems certainly positions it as a strong contender in the rapidly advancing field of agentic AI.

sheikhabrar.aowsaf@gmail.com