How deepauto-ai/automl-agent works
Project type and purpose
• AutoML-Agent is a Python-based multi-agent framework that automates the full AutoML pipeline via LLMs (ICML 2025).
• It orchestrates several specialized agents (Prompt, Data, Model, Operation) under a central AgentManager and interacts with external LLM services (vLLM or OpenAI).
File structure
• Top-level modules:
• agent_manager – core orchestrator (AgentManager, chat loop)
• prompt_agent – handles prompt generation & parsing (WizardLAMP templates, schemas)
• data_agent – inspects data, infers schema, creates preprocessing plans
• model_agent – selects/trains ML models, tunes hyperparameters
• operation_agent – executes code (e.g., training, evaluation)
• prompt_pool – curated prompts per task modality (image, text, tabular, graph, time series)
• experiments – scripts for evaluation, constraint-based or free-form prompting
• Static assets (index.html + static/) suggest a lightweight web interface; example_plans/ provides plan templates.
• configs.py holds LLM backends and API endpoints. requirements.txt pins dependencies.
README for insights
• Describes multi‐agent design, external LLM servers (vLLM or OpenAI), data flow (upload CSV → AgentManager → agents → generated code in agent_workspace).