ScienceDiscovery
中文 GitHub

ScienceDiscovery Core Capabilities

Core does not explain basic Agent concepts. It answers a different question:

What special capabilities does ScienceDiscovery add on top of a general Agent?

If you are new to Agent systems, first read Basic concepts to understand Agent loops, Tools, Skills, Specialists, Workspaces, and Artifacts.

ScienceDiscovery's differentiated capabilities fall into three areas.

1. Exploration and optimization: making research iterative

Idea Tree

For open-ended questions, ScienceDiscovery can explore multiple candidate directions, design approaches, and use feedback to guide further investigation.

RSI for scientific artifacts

For an evaluable artifact, ScienceDiscovery can generate candidates, assess them, and select improved versions through iterative optimization.

Idea Tree asks:

What should we explore next?

RSI asks:

How can the current solution become better?

2. Scientific trust: making results traceable and reviewable

ScienceMemory and Reviewer

Research results need more than generation. They need answers to:

ScienceMemory records relationships across the research process. Reviewer helps identify issues in delivered artifacts.

3. Research execution foundation

These capabilities are the building blocks of an Agent workflow. Their role is introduced in Basic concepts; these pages describe ScienceDiscovery's implementation and usage.

Scientific execution environment and workspaces

Provides code execution, file management, and research environments.

Scientific MCP and Skills

MCP connects external tools and data. Skills provide reusable research methods.

Specialists

Packages responsibilities, methods, and tools into reusable research roles.

These capabilities answer:

How does an Agent perform research tasks?

Idea Tree, RSI, ScienceMemory, and Reviewer answer:

Why is ScienceDiscovery designed for scientific research?

Where to go next