From Code to Docs in 300 Seconds: A Demo of an AI Multi-Agent System

This demo provides a brief, technical deep-dive into DocSmith, a multi-agent system that automates documentation generation from source code. We will live-demo the aigne doc generate command on an undocumented codebase and trace the execution flow through the agentic workflow.

Key technical steps covered include:

• Source Loading: The loadSources agent recursively scans the repository, parsing .gitignore to build a context tree of relevant source files.
• Reflective Structure Planning: A structurePlanning agent proposes a documentation hierarchy, which is then reviewed by a check-structure-planning-result agent in a reflection loop to ensure stability and correctness between runs.
• Content Generation & Validation: A team of contentDetailGenerator agents works in parallel on the approved structure. Each piece of generated markdown is programmatically validated for issues like broken links, malformed tables, and invalid Mermaid diagram syntax before committing the content.
• Git-Based Change Detection: We’ll show how DocSmith uses the delta between the current and previous HEAD commit hashes to intelligently perform incremental builds, only regenerating docs for modified source files.