AutoML on Autopilot

Last Updated on May 4, 2026 by Editorial Team Author(s): Rishav Saigal Originally published on Towards AI. Figure 1 — From a plain-English prompt to a fully tracked MLflow experiment, autonomously. TL;DR Wraps PyCaret’s AutoML engine in a Google ADK agent hierarchy One natural language prompt → plan → code → execution → MLflow tracking Self-corrects up to 10 times on failure; isolates artifacts per session Covers Classification, Regression, Clustering, Anomaly Detection, Time Series If you’ve used PyCaret, you know it already cuts 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. The Root Agent validates your CSV and routes to the right specialist. Each specialist is a SequentialAgent: a Planner designs the pipeline and mints a session ID; an Executor writes the code, runs it, and logs everything to MLflow. Figure 2 — Root routes; each SequentialAgent runs Planner → Executor in strict order. The Smart Bits Session IDs via callback. The Planner outputs a free-text plan with a SESSION_ID: AB1X9Z token. A regex callback extracts it and drops it into shared session state — no structured output format needed. 10-retry self-correction. UnsafeLocalCodeExecutor(error_retry_attempts=10) automatically re-runs generated code on failure, letting the model diagnose and fix its own bugs. Failure short-circuit. A before_model_callback checks a check_failure_status flag and skips re-runs if the task already succeeded — no wasted API calls. Figure 3 — Every metric and param is auto-logged. Named classification_AB1X9Z for instant retrieval. The agent doesn’t just run your ML pipeline — it tracks, isolates, and self-heals through every failure. Run It git clone https://github.com/Rishav1996/PyCaretAgent.gitcd PyCaretAgent && uv pip install .uv run mlflow ui --port 5000uv run adk run pycaretagent Prompt: “Classify heart.csv where the target is ‘target’.” That’s the entire interface. The agent validates the file, plans, codes, executes, and delivers a tracked experiment. Figure 4 — Real-time terminal output. Session ID, retry events, and success signal are all visible in the agent’s log stream. What’s Next This article is the first in a series. Each subsequent piece does a deep-dive into one task type, walking through a real dataset end-to-end — prompt, plan, generated code, and final MLflow results. Figure 5 — Each article in the series covers one task type with a real dataset and annotated agent output. Classification Deep-Dive (Coming Soon) Heart disease prediction with heart.csv. We trace the full agent run — from CSV validation to compare_models() — and annotate every decision the Planner makes. Regression Deep-Dive (Coming Soon) House price prediction. How the Executor tunes via tune_model(), and why the 10-retry mechanism matters when XGBoost hits a dependency mismatch mid-run. Clustering Deep-Dive (Coming Soon) Customer segmentation without a target column. Watch the Root Agent skip target validation entirely and route straight to the unsupervised pipeline. Anomaly Detection Deep-Dive (Coming Soon) Fraud detection on a transactions dataset. The Planner picks Isolation Forest; we break down why, and show how anomaly scores surface as MLflow metrics. Time Series Deep-Dive (Coming Soon) Sales forecasting with seasonality detection. The most complex setup — index parsing, horizon selection, and MASE vs. MAPE in the MLflow comparison table. Future: Deploy Directly to Cloud The current version trains, tracks, and saves models locally. The next major milestone closes the loop — pushing finalized models to cloud storage and inference endpoints using PyCaret’s built-in deploy_model(), triggered directly by the agent with no manual steps. The target UX is a single extra sentence in the user prompt: “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 append a deploy_model() call after finalize_model() — credentials injected from environment variables. A dedicated article in this series will cover the full credential handoff pattern and multi-cloud configuration. 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 all generalize well beyond PyCaret. Github Link : https://github.com/Rishav1996/PyCaretAgent Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI