Yes, Autonomous Kaggle-Solving Agents Already Exist

As of October 18, 2026, the core idea has been built by multiple research groups, including Microsoft, Google and independent open-source developers.

Some systems already do almost exactly what you described: take a Kaggle-style ML problem, generate modeling approaches, train models, evaluate performance, improve their code through repeated experiments and produce solutions.

The opportunity isn’t inventing this category. It’s finding where the existing systems remain inefficient, unreliable, expensive or difficult to operate.

1. The closest existing projects

Microsoft R&D-Agent

Very close match.

Separates research (proposing improvements) from development (implementing them). Supports automated Kaggle feature engineering and model tuning, repeated experimentation and LLM-provider integrations.

GitHub repository · Documentation

The September 2026 release includes additional fine-tuning, evaluation and UI capabilities.

AIDE ML

A direct implementation of the research-and-code-improvement loop. Generates Python solutions, executes them, branches the solution tree, fixes errors and optimizes an evaluation metric.

GitHub repository · Weco commercial platform

Particularly useful as a smaller reference implementation.

Google MLE-STAR

Uses external research to identify candidate models, then performs targeted code refinements and ablations. It addresses a specific weakness of simpler agents: repeatedly rewriting entire training pipelines.

Research paper · Engineering explanation

Google reports medals on approximately 64% of MLE-bench Lite competitions.

MLEvolve

Uses multi-agent collaboration, Monte Carlo graph search and experience-driven memory to explore and improve ML solutions.

GitHub repository

One of the closest matches to the proposed multi-agent experimentation architecture.

Kaggle-AutoML-Agents

A directly competition-focused implementation with Docker sandboxes, session orchestration, feature search, model training and a Kaggle submission workflow.

GitHub repository

Reports a top-6.1% leaderboard placement on Spaceship Titanic as of April 2026. That’s a project-reported result, not broad benchmark proof.

2. How good are these systems already?

The most useful comparison comes from OpenAI’s MLE-bench, which evaluates autonomous agents across 75 Kaggle competitions.

Published MLE-bench leaderboard — percentage of competitions reaching medal-level performance (all splits)

Scores from the published leaderboard; reported runs used different models, time budgets and configurations, so these are not head-to-head tests under equal compute budgets. The leaderboard suspended new submissions on April 24, 2026 while it develops a fairer comparison process, so it should not be interpreted as a current global ranking.

The implication is substantial: this isn’t just a collection of demos. There are already systems with measurable autonomous ML-engineering performance.

But medal rate is not the same as reliability on a new private dataset, production readiness or cost efficiency.

3. Feature comparison against your proposed system

Based on publicly documented capabilities, not independent hands-on testing. A question mark means the feature wasn’t clearly established.

Notice something important: none of these projects needs Firecracker as its defining innovation. Their key engineering contributions are search algorithms, experiment selection, code generation, evaluation and memory.

4. Where is the remaining opportunity?

Here’s my assessment of potential differentiation.

Subjective opportunity scores reflecting differentiation and practical value, not market measurements.

Reproducibility is not just theoretical. An August 2026 Kaggle discussion describes an autonomous agent whose scores varied substantially across eight executions and whose best result could not be reliably reproduced.

And OpenAI’s benchmarking guidance explicitly recommends repeated runs — at least three seeds — because agent performance can have high variance. Its reference evaluation setup allocates up to 24 hours with 36 vCPUs, 440GB RAM and one A10 GPU per task.

That suggests a much more useful optimization target:

Research Efficiency = Verified Model Improvement ÷ (Compute Cost + Inference Cost)

with constraints for reproducibility, stability and safety.

5. What I recommend building instead

I’d avoid starting another autonomous Kaggle framework from scratch.

Build a benchmark-driven experimentation layer on top of an existing agent, initially AIDE ML or Microsoft R&D-Agent.

• Existing research agents — AIDE · R&D-Agent · MLEvolve
• Your experiment orchestrator — scheduling, budget allocation, parallel search
• CPU/GPU workers — containers initially; isolated workers with optional Firecracker
• Immutable evaluator + experiment ledger — scores, reproducibility, cost, failures
• Budget-aware next experiment — feeds back into the orchestrator

The first experiment I would run

Take five MLE-bench Lite competitions. Run AIDE ML and R&D-Agent with identical compute and inference budgets, three seeds each. Preserve their full execution traces and measure results with an independent evaluator.

Then implement just one feature: budget-aware experiment selection, which chooses experiments using expected improvement and estimated compute cost.

Compare against the original agents. If it produces more verified improvement per dollar without hurting reliability, you have a measurable technical contribution. If it doesn’t, you have learned that a more elaborate orchestrator isn’t justified.

Before using competition data, also enforce each competition’s terms around external models, data sharing, internet access and submissions. These rules can differ materially between Kaggle competitions.

Final verdict

For an open-source portfolio, I would prioritize a cost-aware, reproducible autonomous ML research engine, with published benchmarks against AIDE, R&D-Agent and MLEvolve.

For a commercial product, I’d move beyond Kaggle toward autonomous model development and validation for real organizations. Kaggle supplies a convenient benchmark; production ML introduces much harder requirements around data governance, stable performance, deployment and operational risk.

The strongest direction is not another AI that trains models. It is an autonomous research system that can demonstrate, with independently reproducible evidence, that its experiments deliver better results for the resources spent.