EdotEnv

AI agents that are meant to recursively self-improve need a steady supply of problems that keep getting harder, otherwise the improvement loop stalls. Frontier AI labs and academic groups building research harnesses, evaluation benchmarks, and post-training environments need a source of tasks with this escalating difficulty built in. Static benchmarks do not naturally generate harder follow-on problems once an agent solves them. Financial markets have this property inherently: every useful discovery attracts competition, and every successful strategy makes the next opportunity harder to find, which the text frames as a natural proving ground that current environments lack.

EdotEnv builds RL environments at scale from financial market data that increase in difficulty so agents can continuously hillclimb. These are multi-step environments in which agents form hypotheses, design experiments, verify results, and iterate, mirroring a research loop rather than a single-shot task. EdotEnv uses these environments to evaluate and post-train agents on increasingly difficult research tasks, and works with frontier AI labs and academic groups on research harnesses, evaluation benchmarks, and post-training environments.