How vidyutchakrabarti/agentflux works
Project type and purpose
• AgentFlux is a full-stack AI agent engineering platform. The Next.js/React frontend (in main/app and main/components) provides an interactive playground, visualization, and observation UI. The Node/TypeScript backend (in main/server.ts, API routes, and the main/Agents folder) handles agent orchestration, refinement, persistence, and third-party integrations.
File structure
• Top-level directories:
• data – static training and lookup data (codes.json, train.csv)
• main/Agents – core “agent” code: two ADK agents (graph_architect, prompt_refiner) plus shared helper_funcs and api wrappers
• main/app – Next.js 13 app folder with pages, API routes, and Firebase auth
• main/components – reusable UI components for visualization, code editor, modals, etc.
• main/lib, main/models – MongoDB connection (mongodb.js) and Mongoose schemas (Playground, Sheet, Runs, RefineHistory)
• main/public – static assets (AgentFlux-logo.gif, videos)
• Configuration: .env.example, package.json, tsconfig, ESLint, Tailwind
README for additional insights
• Confirms the UI features (playground, graph viz, diff reports), the two core agents (Prompt Refinement, Graph Re-architect), model management (Hugging Face models behind scenes), and observation tooling (logs, profilers).
• Mentions EC2-backed “studios” but that detail can be captured later as an optional compute layer.