Manus test and related papers

Written by

Iris Vance

Updated on:July-13th-2025

Manus is an AI agent company founded by Chinese serial entrepreneur Xiao Hong (post-90s, alumnus of Huazhong University of Science and Technology), affiliated to Butterfly Effect PTE. LTD. Its core product is the world's first general-purpose AI agent, which integrates multi-model capabilities through a multi-agent architecture to achieve autonomous task planning and execution (such as resume screening, stock analysis), and surpassed OpenAI's similar products in the GAIA benchmark test..

Relying on the tens of millions of overseas users accumulated by the early Monica plug-in, the team focused on "delivering results" rather than "generating answers" and accurately entered into enterprise efficiency scenarios. However, the official website was in English and the technology relied on model encapsulation, which caused controversy over its "domestic nature". The internal test code was hyped to tens of thousands of yuan, reflecting the coexistence of market enthusiasm and doubts .

Measured example

Multi-agent collaboration and task decomposition

The core capability of Manus lies in the collaborative work of multiple agents, which involves the research of task decomposition and distributed execution . Related papers may include:

Hierarchical Task Network Planning for Multi- Agent Systems
Cooperative Multi-Agent Reinforcement Learning
Tool Learning with Foundation Models, tool learning based on foundation models. This type of research may support Manus's ability to call tools such as browsers and code editors .

Automated tool invocation and end-to-end execution

Manus is unique in that it directly performs complex tasks (such as code generation and data analysis), which may involve the following directions

AutoGPT: Autonomous Task Automation with Large Language Models, automated task execution based on large language models.
Code Generation and Execution via LLMs, code generation and execution based on large models, such as OpenAI Codex or DeepSeek Coder.
Chain-of-Thought Prompting for Complex Task Solving, chain-of-thought prompting technology, used for task planning and step explanation.

Human-machine collaboration and explainable AI

Manus emphasizes "human-machine collaboration" and transparent operation. Related research may include:

Human-AI Collaboration in Task-Oriented Systems Human-AI Collaboration in Task-Oriented Systems.
Explainable AI for Autonomous Agents is a study on the explainability of autonomous agents, used for progress feedback and logical description of Manus .

Virtual Reality and Motion Capture Technology

Manus's virtual reality gloves, such as the Quantum Meta Glove, are relevant to the following areas:
High-Precision Magnetic Tracking for VR Gloves VR, high-precision magnetic tracking technology for gloves.
Real-Time Motion Capture in Robotics Training, real-time motion capture in robot training), such as Tesla's use of Manus gloves to train Optimus robots .

Benchmarking and performance evaluation

Manus achieves SOTA performance on the GAIA benchmark. Related papers may include:

GAIA: A Benchmark for General AI Assistants, General AI Assistant Benchmark.
Evaluating Autonomous Agents in Complex Environments, Evaluation of autonomous agents in complex environments.

AI to control and execute tasks