Leeroo open-sources Kapso, an engine that learns from experiments on enterprise data

Leeroo has open-sourced Kapso, an engine that takes an objective and success criteria, runs experiments in a sandbox, and works toward a solution that can be deployed to the user's stack. The team says it is aimed at large, messy data estates where investigation, hypotheses and testing have to happen before anything reaches production.

Kapso keeps a bank of lessons, each with a credence score, conditions on where it applies and the evidence behind it. An ideation module proposes ideas informed by those lessons and ranks them by expected return against budget, while a separate verifier module checks results that experiments cannot access. Banks can be A/B tested and shared through git using the kapso experience commands.

The team reports first place on RelBench predictive modelling against the best tabular foundation model, first place on OpenAI's MLE-bench at submission time, and a top-three finish among AI systems at the 2026 International Olympiad in AI, ahead of 471 human contestants. The kapso.deploy call runs a shorter loop to adapt a solution to targets such as Docker, Modal, BentoCloud or LangGraph.

Developers building agents that tune models or pipelines against measurable criteria can look at an open-source loop that stores what it learns in a git-backed bank instead of in model weights. The benchmark numbers are self-reported, and a commenter pointed to RelArena as a more controlled comparison, so check results before relying on them. The project says it is model-agnostic and coding-agent-agnostic, so it can be tried with different coding agents.