Internal Research Agent Architecture
Build internal research agents for long analysis, artifacts, approval, and source tracking.
Key takeaways
- Internal research agents work best when they produce inspectable artifacts, preserving sources, assumptions, intermediate steps, and final recommendations.
- The design pattern moves through brief, plan, collect, analyze, artifact, and review stages.
- Use Workflow for long-running research and Sandbox for file generation or code analysis.
- Use MCP for permission-aware internal sources and Gateway routing for cost and quality classes.
- Store artifacts with stable IDs and source references so outputs stay traceable.
Internal research agents are best when they produce inspectable artifacts, not just long answers. They should preserve sources, assumptions, intermediate steps, and final recommendations.
Design Pattern
| Stage | Output |
|---|---|
| Brief | Question, scope, sources, deadline |
| Plan | Research steps and assumptions |
| Collect | Source notes and extracted facts |
| Analyze | Comparison, synthesis, risks |
| Artifact | Memo, table, deck outline, or dataset |
| Review | Human comments and follow-up tasks |
Runtime Choices
- Use Workflow for long-running research.
- Use Sandbox for file generation or code analysis.
- Use MCP for permission-aware internal sources.
- Use Gateway routing for cost and quality classes.
- Store artifacts with stable IDs and source references.