EdotEnv

EdotEnv builds evals and trains AI agents on quantitative research tasks. Its thesis is that agents should learn how to do work, not just what to do in specific tasks. As LLMs improve, static data saturates and become less meaningful for model training and comparison. Useful training tasks should increase in difficulty as models advance. EdotEnv turns trading based data-science problems into complex, multi-step research environments.

Agents iterate over hypotheses forming, experiment choosing and running, tool use, result evaluation and resource allocation over multiple turns. The same environments are used for both evaluation and post training. Trading in the market is a naturally moving benchmark: as traders profit from inefficiencies, the market becomes more efficient and inefficiencies become harder to find. This way the market provides LLM training with a natural, non saturating testbed.