ml-agents · private early access

Machine learning that runs itself.
And audits itself.

ml-agents turns a parquet file into a stakeholder-ready model. Point it at your data, answer a short interview, and a supervised agent harness runs the feature-engineering loop on your own machine — sandboxed, audited, and reported. Built on Claude Code.

3task types
11optimization metrics
126automated tests
100%on your hardware
How it works

Two agents. One climbs, one supervises.

Unsupervised automation is how AutoML overfits, leaks labels, and burns compute. ml-agents splits the work between two roles with an enforced boundary.

Outer supervisor

Watches everything. Trusts nothing.

Your main Claude Code session scaffolds the project, then audits on a loop — overfit gap, leakage smell, progress, liveness, memory — and decides when to stop. At the end, it writes the final report.

Inner hill-climber

Experiments relentlessly. Inside a sandbox.

A background agent confined to its workspace edits features.py, model.py and analysis.py, runs experiments, journals every result, and reverts rejected experiments via git. It can't read or write anything outside its walls.

Features

Everything a careful ML engineer would do. Automatically.

Guided interview

Infers your task from the target column, proposes time-based splits from a label-availability diagnostic, and flags leakage-prone columns — before a single experiment runs.

Supervised hill-climbing

The inner agent iterates features and models toward your metric. Every experiment is journaled; every rejected one is reverted via git. Nothing is lost, nothing lingers.

Continuous audits

Overfit gap, leakage smell, progress and health, liveness, memory pressure, and a principled stopping decision — checked on a loop, not after the damage is done.

Sandboxed by design

Hook-based guardrails confine the inner agent to its workspace — enforced by agent identity, so it can't escape by changing directories. No containers, no cluster, no infra.

Time-aware validation

Expanding-window splits with outer test windows that stay silent during development and are only inspected at report time. The test set can't leak into your iteration loop.

Two-tier reporting

An operator debug log for you, and a stakeholder model card for risk, compliance, and product — with real feature importance, SHAP, plots, and insights. Not placeholders.

Scope

Any tabular use case. The metric you actually care about.

A metric registry carries each metric's direction, so hill-climbing on rmse means going down — and the audit, reporting, and stopping logic all agree.

Binary classification

giniauclog_lossaccuracyf1

Regression

rmsemaer2

Multiclass classification

macro_f1accuracylog_loss
EnginePolars + LightGBM
Specone automl.yaml drives everything
Escape hatchescustom loaders & splitters
Runs onmacOS & Linux, with Claude Code
Early access

ml-agents is in private early access.

We're onboarding a small group of data science teams. Tell us what you're modeling and we'll be in touch.

Prefer email? Write to trantrikien239@gmail.com.