AgentKaggle
Real competitions, real leaderboards, and an audit that refuses to flatter itself. An official final rank and a late-submission estimate are two different claims, and they are never reported as one.
github.com/agentkaggle · the leaderboard · the audit
What the two words mean
This is the part that matters, and the part most write-ups skip.
Official means the run was entered before the deadline and holds an exact final position on the Kaggle leaderboard. Of the thirteen completed competitions, the best official finishes are 67 of 3,355 on Student Health Risk and 137 of 6,125 on ROGII — both inside the top 2.5%.
Late means only the score is real. The submission was made after the competition closed, scored by Kaggle, and then placed against the frozen final board to estimate where it would have landed. Several of these are at or near the top — 1 of 1,142 on Google Code Golf 2025, 1 of 670 on WiDS Datathon 2024, 5 of 2,657 on CMI Sensor — and none of them is a rank, a medal, or evidence of having competed. They are an estimate of score strength and are labelled as one everywhere they appear.
Mixing the two would make a much better headline. It would also be the exact failure mode our flows are built to catch, so we do not.
The audit
Every number above comes from agentkaggle/kaggle-results-audit, which is a repository rather than a claim:
- Completed results — thirteen competitions, forty-nine person-competition results, each classified official or late, each mapped to the session or supporting evidence behind it.
- Ongoing results — nine competitions, twenty-five results, captured from official public leaderboards with the timestamps recorded. Current public ranks are volatile and say so.
- Coverage and limitations — scan coverage, failure counts and truncation are reported, so a partial result is never dressed up as a complete one. Ten tracked entrants had no valid result in the completed set, and that is written down too.
The live leaderboard is regenerated from the Kaggle API and shows public and private rank and score separately, with the medal zone only estimated where Kaggle says the competition awards points.
Why Kaggle
A Kaggle competition is a long-horizon task with an unarguable score, a deadline, and thousands of humans working the same problem. It is one of the few settings where an agent loop can be wrong in public.
It also fails in the way we care about most. A model left alone will narrow the validation split until the number improves, or fit the public leaderboard and fall off the private one. The private board is the check, and it arrives after everything is already decided.
What is public, and what is not
The audit and the leaderboard are public repositories, and they are where every number on this page comes from.
The per-entrant repositories of workflows and sessions are not public. They hold competition data and account credentials, and several of the competitions are still running. Rather than link you to a page you cannot open, the audit carries what can be published: for each best result, the session or supporting evidence it was traced to, and an explicit note where a result could not be traced to one.
The loops themselves are the flows we write about elsewhere, run on Humanize 2.