What Are the 5 Specialized Agents in the /understand Command?

The /understand command orchestrates five specialized agents—Project-Scanner, File-Analyzer, Architecture-Analyzer, Graph-Reviewer, and Assemble-Reviewer—that sequentially discover, parse, architect, validate, and assemble a unified knowledge graph of your codebase.

In the Lum1104/Understand-Anything repository, the /understand slash command is the primary entry point for automated codebase analysis. The command delegates every stage of its workflow to one of five specialized agents defined in the understand-anything-plugin/agents/ directory. This modular architecture keeps the pipeline testable and easy to extend.

The 5 Specialized Agents in the Pipeline

The analysis workflow is divided into five distinct stages. Each stage is owned by a single agent with a narrowly defined responsibility.

1. Project-Scanner

The Project-Scanner agent initiates the pipeline by walking the file system, discovering the project root, and building an initial list of files to process. Its configuration and behavioral spec live in understand-anything-plugin/agents/project-scanner.md.

2. File-Analyzer

The File-Analyzer agent consumes that file list and runs the bundled tree-sitter extraction script—or language-specific parsers—to produce deterministic, high-quality structural data for each source file. You can review its parsing rules in understand-anything-plugin/agents/file-analyzer.md.

3. Architecture-Analyzer

The Architecture-Analyzer agent ingests the structural data and, with LLM assistance, infers a high-level architecture diagram including components, dependencies, and data flows. Its logic is documented in understand-anything-plugin/agents/architecture-analyzer.md.

4. Graph-Reviewer

The Graph-Reviewer agent validates the generated knowledge graph against the schema, flags inconsistencies, and suggests fixes before the graph is finalized. Its validation criteria are stored in understand-anything-plugin/agents/graph-reviewer.md.

5. Assemble-Reviewer

The Assemble-Reviewer agent—sometimes referred to as the "reviewer" in internal docs—pulls together the final human-readable report. It merges insights from the other agents and writes the output that the /understand command ultimately returns. Its assembly logic is defined in understand-anything-plugin/agents/assemble-reviewer.md.

How the Agents Are Orchestrated

These agents are not executed in isolation. According to the Understand-Anything source code, the command handler in understand-anything-plugin/src/understand-chat.ts sequentially dispatches each agent and aggregates the intermediate results into a single knowledge graph. This central orchestrator enforces ordering, passes context between stages, and surfaces the final report to the user.

Extending the Pipeline: The /understand-domain Variant

The default pipeline uses the five agents above, but the repository supports an optional sixth agent called Domain-Analyzer. When you invoke the /understand-domain variant, this agent enriches the graph with domain-specific terminology and ontology, effectively turning the pipeline into a six-agent workflow.

Practical Code Examples

Below are concise snippets showing how to invoke the command and customize the pipeline.

// Example 1 – Basic usage (runs the default 5 agents)
await clerk.runCommand('/understand', {
  // No extra options – the command auto-discovers the project and
  // runs Project-Scanner → File-Analyzer → Architecture-Analyzer → Graph-Reviewer → Assemble-Reviewer
});

// Example 2 – Adding the optional Domain-Analyzer
await clerk.runCommand('/understand-domain', {
  // The extra "domain-analyzer" agent enriches the graph with domain-specific
  // terminology and ontology.
});

// Example 3 – Running a single agent manually (for debugging)
await clerk.runAgent('file-analyzer', {
  targetFiles: ['src/**/*.ts'],
});

Summary

• The /understand command relies on five specialized agents to analyze a codebase end-to-end.
• Project-Scanner discovers files, File-Analyzer extracts structure, Architecture-Analyzer infers design, Graph-Reviewer validates the graph, and Assemble-Reviewer produces the final report.
• Orchestration is centralized in understand-anything-plugin/src/understand-chat.ts.
• Each agent spec lives in a dedicated markdown file under understand-anything-plugin/agents/.
• The pipeline can be extended to six agents via the /understand-domain variant and the Domain-Analyzer.

Frequently Asked Questions

What is the role of the Project-Scanner agent?

The Project-Scanner agent is responsible for the discovery phase. It walks the file system, identifies the project root, and compiles the initial list of source files that downstream agents will analyze.

How does the File-Analyzer agent parse source files?

The File-Analyzer agent runs a bundled tree-sitter extraction script—or alternative language-specific parsers—against each discovered file to generate deterministic structural data that the Architecture-Analyzer consumes.

What is the difference between Graph-Reviewer and Assemble-Reviewer?

The Graph-Reviewer agent validates the knowledge graph against its schema and flags logical inconsistencies, while the Assemble-Reviewer agent focuses on presentation: it merges all insights into the final human-readable report returned by the /understand command.

Can I extend the /understand pipeline with additional agents?

Yes. The repository provides a /understand-domain variant that appends a sixth Domain-Analyzer agent to the default pipeline, enriching the output with domain-specific terminology and ontology.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}