Using Google Antigravity to Build an Autonomous ML Swarm

What happens when you use an AI agent system to build another AI agent system?

I tried this and created an ML Agent Orchestrator—an autonomous multi-agent swarm designed to handle end-to-end data science workflows, from exploratory data analysis (EDA) and feature engineering to model training and explainability.

How It Works

Using Google Antigravity as the core agentic tool, I built a specialized backend swarm where dedicated agents hand off state, execute code, and synthesize insights in real time.

Key Features & Capabilities

1. Multi-Agent Swarm Orchestration Engine

• Master Director Agent (orchestrator.py): Coordinates execution topology, maintains shared workspace context, handles inter-agent dependencies, and yields real-time Server-Sent Events (SSE).
• EDA & Data Quality Agent (eda_agent.py): Profiles dataset shape, computes numeric/categorical distribution statistics, correlation matrices, and identifies data quality anomalies.
• Feature Engineering Agent (feature_agent.py): Imputes missing values, log-transforms skewed distributions, one-hot encodes categorical variables, applies standard scaling, and expands the feature space.
• AutoML & Model Trainer Agent (automl_agent.py): Fits and cross-validates multiple candidate models (Random Forest, Gradient Boosting, Ridge/Logistic Regression, Decision Trees), building a leaderboard to select the champion architecture.
• Explainability & Metrics Agent (explainability_agent.py): Computes feature importances, confusion matrices, residual error metrics, and plain-English diagnostic drivers.
• Report & Code Generator Agent (reporter_agent.py): Synthesizes the run into an executive Markdown report and generates executable, standalone Python code (pipeline.py).

2. Modern Interactive Web Dashboard (React + Vite + Tailwind CSS)

• SVG DAG Execution Graph: Live visual node graph showing node status (Pending, Running, Succeeded), timing metrics, pulse animations, and clickable node output inspector modals.
• Inter-Agent Live Message Console: Real-time streaming log of agent reasoning, state transfers, and tool events.
• Dataset Playground: Pre-loaded benchmark datasets (Telecom Churn, House Prices, Credit Risk Default) plus drag-and-drop CSV file upload with data preview table and target column picker.
• Model Insights & Code Export: Leaderboard comparison table, feature importance bar visualizer, executive Markdown report viewer, and one-click Python pipeline code download.

Dual Frontend Experience

To make the swarm accessible across different setups, I built two ways to interact with it:

1. Rich Local GUI: A custom-built local React dashboard featuring a real-time DAG execution graph, live inter-agent message logs, and deep metric insights.

2. Global Hugging Face Space: A lightweight, pure-Python Gradio interface hosted on Hugging Face Spaces for instant online access without local setup.

Try out the live web app:
Hugging Face Spaces - ML Agent Orchestrator