AutoML on Autopilot | towards AI

Author(s): Rishav Sehgal

Originally published on Towards AI.

Figure 1 – From a plain-English prompt to a fully tracked MLflow experiment, autonomously.

TL;DR

Google wraps PyCaret’s AutoML engine in an ADK agent hierarchy
A natural language prompt → plan → code → execution → MLflow tracking
Self-correction up to 10 times upon failure; isolates artifacts per session
Includes classification, regression, clustering, anomaly detection, time series

If you’ve used PyCaret, you know that it already cuts out ML boilerplate dramatically. PyCaretAgent goes further: a root agent reads your intent, a planner designs the pipeline, and an executor writes and runs the code – all without you touching a line of Python.

how it works

Three layers. inert agent Validates your CSV and routes it to the right expert. Every expert is one SequentialAgent: A planner Designs the pipeline and creates a session ID; One Executor Writes code, runs it, and logs everything to MLflow.

Figure 2 – Basic route; Each SequentialAgent runs Planner → Executor in a strict order.

smart bits

Session ID via callback. Planner outputs a free-text plan SESSION_ID: AB1X9Z Token. A regex callback extracts it and leaves it in the shared session state – no structured output format needed.

10-Try self-improvement again. UnsafeLocalCodeExecutor(error_retry_attempts=10) Automatically re-runs the generated code on failure, allowing the model to diagnose and fix its own bugs.

Failure short-circuit. A before_model_callback checks a check_failure_status Flags and skips re-runs if the task has already succeeded – no wasted API calls.

Figure 3 – Every metric and parameter is auto-logged. nominated `classification_AB1X9Z` For immediate recovery.

The agent doesn’t just run your ML pipeline – it tracks, isolates, and fixes itself every failure.

run it

git clone https://github.com/Rishav1996/PyCaretAgent.git
cd PyCaretAgent && uv pip install .
uv run mlflow ui --port 5000
uv run adk run pycaretagent

Hint: “Classify heart.csv where target is ‘target’.” That’s the entire interface. The agent verifies the file, plans, codes, executes, and delivers a tracked experiment.

Figure 4 – Real-time terminal output. The session ID, retry events, and success signals all appear in the agent’s log stream.

what will happen next

This article is the first in this series. Each next piece dives deeper into a task type, walking through a real dataset from start to finish – prompt, plan, generated code and the final MLFlow result.

Figure 5 – Each article in the series covers a task type with real datasets and annotated agent outputs.

Classification Deep-Dive (Coming Soon)

with a predisposition to heart disease heart.csv. We trace the entire agent run – from CSV validation to compare_models() – and explain each decision taken by the planner.

Regression Deep-Dive (coming soon)

House price prediction. How the executor plays the tune tune_model()And why the 10-retry mechanism matters when XGBoost hits a dependency mismatch in mid-run.

Clustering Deep-Dive (coming soon)

Customer segmentation without target column. Notice that the route agent skips target validation altogether and goes straight to the unsupervised pipeline.

Deep-dive into anomaly detection (coming soon)

Fraud detection on transaction datasets. The planner chooses the isolation forest; We explain why, and show how anomaly scores are revealed in the form of MLflow metrics.

Time Series Deep-Dive (Coming Soon)

Sales forecasting with seasonality identification. Most complex setup – index parsing, horizon selection and MASE vs MAPE in MLflow comparison table.

Future: Deploy directly to the cloud

The current version trains, tracks, and saves models locally. The next major milestone closes the loop – pushing the final model to cloud storage and using PyCaret’s built-in inference endpoint. deploy_model()Triggered directly by the agent without any manual steps.

The target UX user prompt has an additional sentence: “Classify heart.csv, target=’target’, deploy to AWS.” The root agent will parse the platform, pass it as a session state variable, and the executor will add a deploy_model() call later finalize_model() – Credentials injected from environment variables. A dedicated article in this series will cover full credential handoff patterns and multi-cloud configurations.

PyCaretAgent is a clean, reusable template for any agent-wrapped AutoML system. The planner/executor pattern, state handoff via callbacks, and retry-based self-correction are all generalized far beyond PyCaret.

Github link: https://github.com/Rishav1996/PyCaretAgent

Published via Towards AI

AutoML Autopilot