AgenticML

Multi-agent ML pipeline — automated machine learning powered by LangGraph and LLMs.

AgenticML assembles a team of specialised AI agents (Planner, Profiler, Cleaner, Featurizer, Modeler, Evaluator, Critic, Reporter) that collaborate iteratively to build, evaluate, and refine ML models on your data.

Installation

pip install swayamml

Optional providers (only install the one you use):

pip install swayamml[anthropic]   # Claude models
pip install swayamml[google]      # Gemini models
pip install swayamml[boost]       # XGBoost + LightGBM
pip install swayamml[all]         # Everything

Quick Start

Python API

from agenticml import ml

# Minimal — target and problem type are auto-detected
result = ml.run("data.csv")

# Explicit options
result = ml.run(
    "data.csv",
    target="price",
    problem_type="regression",
    metric="rmse",
    model="gpt-4o",          # or "claude-3-sonnet-20240229", "gemini-pro"
    api_key="sk-...",         # optional — falls back to env var
    verbose=True,             # print LLM prompts & responses
    max_iterations=3,
)

CLI

# Uses OPENAI_API_KEY from environment
agenticml --file data.csv --target price --verbose

# Specify model and key
agenticml --file data.csv --model claude-3-sonnet-20240229 --api-key sk-ant-...

# All options
agenticml --file data.csv \
    --target label \
    --problem_type classification \
    --metric f1 \
    --model gpt-4o \
    --max_iterations 3 \
    --verbose \
    --stream

You can also run via module:

python -m agenticml --file data.csv

LLM Provider Support

The provider is auto-detected from the model name:

Model prefix	Provider	Env variable	Install extra
gpt-*, o1*, o3*	OpenAI	OPENAI_API_KEY	(included)
claude-*	Anthropic	ANTHROPIC_API_KEY	pip install swayamml[anthropic]
gemini-*	Google	GOOGLE_API_KEY	pip install swayamml[google]

Pass the key directly or set the environment variable:

export OPENAI_API_KEY=sk-...

Pipeline Architecture

Planner → Profiler → Cleaner → Featurizer → Modeler → Evaluator → Critic
                                                                      ↓
                                                              (blocking issues?)
                                                              ↓ yes        ↓ no
                                                          Orchestrator   Reporter
                                                              ↓
                                                          (next iteration)

Each run produces:

• report.md — human-readable summary
• run_manifest.json — full reproducibility metadata
• Trained models, evaluation plots, and intermediate data in the runs/ directory

Verbose Mode

Use --verbose (CLI) or verbose=True (Python) to see exactly what each agent sends to the LLM and what it gets back — useful for debugging and understanding pipeline decisions.

License

MIT