Numel Playground: The Next-Generation Orchestration Platform for Visualizing AI Agent Workflows

Numel Playground is an autonomous AI agent workflow visual editor that combines a node-based graphical canvas with a Python backend. It supports automatic workflow generation, self-optimization evaluation loops, and unified multi-platform integration, providing a brand-new visual orchestration paradigm for AI application development. It addresses the pain points of traditional development, such as excessive boilerplate code, and the lack of flexibility in existing low-code platforms.

Traditional AI application development requires a lot of boilerplate code, while existing low-code platforms lack flexibility and intelligent capabilities. Numel is positioned as the "ComfyUI for AI Agents", aiming to make any AI-driven automation process visual and orchestratable. Its core slogan is "ComfyUI generates images, Numel generates optimal results—automatically", emphasizing its intelligent self-optimization capability.

Backend

Based on the FastAPI framework, node types are defined using Pydantic. The Python schema source code is directly sent to the frontend to dynamically build the node panel. Core components include the agent backend (manages lifecycle), workflow engine (executes processes), and evaluation & planner (automatically optimizes processes).

Frontend

The Canvas graphic editor is implemented with native JavaScript, pre-packaged with dependencies like CodeMirror. The frontend and backend communicate via REST APIs, and real-time events (execution progress, streaming output) are transmitted via WebSocket.

1. Dynamic UI: The frontend interface is automatically generated based on the backend Pydantic schema; no frontend development is needed for adding new nodes.
2. Text-Driven Generation: Automatically generate node graphs via natural language descriptions.
3. Self-Optimization Loop: The eval_flow component works with the planner to implement A/B testing, feedback collection, and process adjustments.
4. Real-Time Browser ML: Integrates MediaPipe to support pose/facial/gesture recognition.
5. Multi-Channel Integration: Supports 9 platforms + Web console, with multi-tenant architecture to isolate user spaces.
6. Background Agent Tasks: Scheduled/event-driven proactive tasks (monitoring, data collection, etc.)