Building MCP-native hierarchical AI scientist ecosystems: a perspective on scaling multi-agent scientific discovery

Area of Science:

• Artificial Intelligence
• Scientific Research
• Computational Science

Background:

• Large language models (LLMs) are transitioning from basic chatbots to sophisticated agentic AI systems capable of complex research tasks.
• Current LLMs face limitations in real-world scientific impact due to poor interoperability and challenges in scalable coordination among multiple agents.
• The existing paradigm of a single 'AI scientist' is insufficient for broader scientific application.

Purpose of the Study:

• To propose a new framework for agentic scientific discovery that addresses current LLM limitations.
• To advocate for an ecosystem of protocol-native agents and tools organized hierarchically.
• To outline pathways for scaling AI-driven scientific discovery through enhanced interoperability and coordination.

Main Methods:

• The paper presents a perspective on the evolution of LLMs in scientific research.
• It introduces the Model Context Protocol (MCP) as a key component for interoperability.
• It proposes three strategies for scaling MCP-native scientific ecosystems: expert-maintained MCP servers, automated code transformation, and autonomous agent evolution.

Main Results:

• The development of an ecosystem of protocol-native agents and tools is crucial for the next phase of AI-driven scientific discovery.
• MCP serves as an emerging standard for scientific tool and context exchange, enhancing interoperability.
• Three pathways are identified to improve the scalability and composability of scientific AI ecosystems.

Conclusions:

• Moving beyond monolithic AI scientists to a hierarchical ecosystem of agents and tools is essential for unlocking the full potential of LLMs in science.
• Implementing MCP and the proposed scaling strategies will accelerate AI-driven scientific discovery.
• A practical roadmap is provided for expanding tool supply and coordination within these ecosystems.