Concepts: how ScienceDiscovery completes a research task

Before using ScienceDiscovery, understand a few basic concepts. The Core capabilities section then explains the research-specific capabilities built on top of a general Agent system.

How a task runs

A ScienceDiscovery task can be viewed as:

Research question
      ↓
Research Agent
      ↓
Understand goal, plan steps, select capabilities
      ↓
Call tools, execute code, collaborate with roles
      ↓
Create research Artifacts
      ↓
Review, reuse, and continue research

The Agent does not only generate text. It works toward a goal by understanding the problem, selecting methods, executing actions, and delivering results.

The Agent Loop

A typical Agent run includes:

Understand the objective and current context;
Decide what information or action is needed next;
Call tools, execute code, or ask other roles for help;
Adjust based on execution results;
Deliver an answer and research artifacts.

The execution timeline shows what actually happened during a task.

What capabilities can an Agent use?

Three extension concepts are easy to confuse:

Concept	Purpose	Simple view
MCP	Provides external tools and data interfaces	Agent tools
Skill	Provides reusable methods and workflows	Agent methods
Specialist	Provides focused responsibilities and roles	Agent roles

For example:

MCP can let an Agent query a literature database;
a Skill can guide a literature-review workflow;
a Specialist can define a dedicated research role.

Workspace, Artifacts, and research delivery

Files created during research have different purposes.

Workspace
  ├── Uploaded data
  ├── Temporary code
  ├── Intermediate results
  ↓
Artifact
  ├── Reports
  ├── Code
  ├── Tables
  └── Images

A useful shorthand:

Workspace is the workbench; Artifact is the deliverable.

Not every file needs to become an Artifact. Results worth inspecting, downloading, reusing, or continuing should be delivered as Artifacts.

Execution environment

The Agent needs a place to perform real work.

ScienceDiscovery provides a research execution environment where the Agent can:

run Python, R, and Shell;
analyze data;
preserve scripts and results;
execute computation inside a controlled environment.

The execution environment determines where work happens; Artifacts determine what remains as a result.

From basic concepts to research capabilities

After understanding these foundations, continue with:

Core capabilities: why ScienceDiscovery is designed for research workflows.
Domain guides: see complete workflows through real tasks.