EdotEnv is a Y Combinator-backed platform that provides market-derived reinforcement learning environments to teach LLMs and agents applied machine learning and long-horizon planning under adversarial noise.
We are Rui and Michael and we’re building EdotEnv (<a href="https://edotenv.com" rel="nofollow">https://edotenv.com</a>): self-improving RL environments from Quant Trading workflows.<p>With all the benchmaxxing around, evals saturate and become meaningless for model comparison. Useful benchmarks should increase in difficulty as models advance. Back in our Quant jobs, Michael and I saw that the market has exactly this property: markets became more efficient as people profited from trading inefficiencies, making new profitable strategies harder to find and old ones decay over time.<p>This makes markets an ideal, continuously evolving benchmark for LLM training. The hard part is to turn professional quant workflows into reliable training envs, as this is a very niche expertise.<p>In our environments, we give LLMs a quant trading workflow and evaluate their performance on out-of-sample data: build predictive features/ models, design a portfolio, backtest strategies, adapt continuously to market regimes. Each step is a task with different self-built tools. For example, a predictive feature building task gives the agent cleaned market data of time period [0,T] to research ideas, a backtesting tool to test created features at time t on [0, t], an execution tool to trade strategies with the new features on [t+1, T] and a final evaluation. Our reward isolates the agent's feature building skills and yet benefits from market properties.<p>From running SOTA models in our environments, we see that i) they seem to struggle with iterating deeply on research ideas, preferring broad shallow searches; ii) higher reasoning does not seem to increase performance and iii) agents do not understand trading, e.g. when losing money they stop trading instead of trading smarter. Check out our blogs for more details! <a href="https://edotenv.com/?tab=blog" rel="nofollow">https://edotenv.com/?tab=blog</a><p>Quant workflows are essentially applied ML research, long-horizon planning and continual learning. Through our envs, we teach these transferable research skills, rather than task specific answers. Our environments are closer to a realistic research workflow: we use real-world data instead of synthetic ones; our envs naturally contain noise and real trade-offs; our rewards are verifiable and immediate, with no need for an additional LLM judge or human expert.<p>We open sourced a sample task repository: <a href="https://github.com/MMcollab-dotcom/feature-engineering" rel="nofollow">https://github.com/MMcollab-dotcom/feature-engineering</a>. We plan to sell continuously improving envs to AI labs/researchers/enterprises training their own agents, who are interested in ML modelling capabilities, continual learning, long horizon planning or Quant Research in general.<p>We'd love feedback from anyone trying out their own agents in our envs, for either eval or post training. And of course, we are always happy to discuss the future of trading with LLMs (and no, it should not be asking the LLM to read tea leaves and give you the stock to buy tomorrow). Looking forward to your comments!
# E.env — Environments for intelligence that adapts
Source: [https://edotenv.com/](https://edotenv.com/)
Reinforcement learning environments
## We build market\-derived RL environments to teach agents applied ML and long\-horizon planning under adversarial noise\. Static synthetic benchmarks are unrealistic and saturate quickly, whereas markets are non\-saturating and self\-improving\.
[](https://www.ycombinator.com/)
THESIS
Static worlds produce static intelligence
## Quant is the hardest, yet solveable data science task\.
We programmatically generate quant research tasks inside environments built from real market data\. Agents use professional tools—and build their own in Bash—to make trading decisions and develop profitable strategies\.
Markets do not saturate: successful trading makes them more efficient, while edges decay and regimes shift\. That makes our environments a continuously harder benchmark for improving models\.
HORIZON
A decision is not a moment
## Trading decisions affect more than one future step\.
Trading successfully means planning ahead multiple steps and assess trade\-offs between short and longterm gains
T\+00Choose
### Act under partial information\.
The model sees an incomplete state and commits before the full consequences are observable\.
T\+18HCompound
### The decision becomes part of the environment\.
Exposure, opportunity cost and every action not taken reshape the path that follows\.
T\+53HRevalue
### The objective moves\.
A decision can remain locally correct while becoming globally expensive as conditions drift\.
T\+96HAdapt
### The policy that worked has expired\.
Success belongs to the model that recognizes the new regime before yesterday’s behavior becomes consensus\.
This paper introduces LLM-as-Environment-Engineer, a framework where LLMs design their own training environments for reinforcement learning in multi-agent reasoning tasks, enabling self-improving training that surpasses larger proprietary models.
EnvFactory automates the creation of executable tool environments and natural multi-turn trajectories for training LLMs with agentic reinforcement learning, achieving superior performance on benchmarks like BFCLv3 and MCP-Atlas with fewer environments than prior work.
OpenEnv, a platform for reinforcement learning environments, is expanding its tutorials, covering topics like evaluating agents, rewards via rubrics, and connecting agents via MCP.
EnvScaler is an automated framework for scaling tool-interactive environments for LLM agents through programmatic synthesis, creating 191 diverse environments and 7K scenarios to improve agent performance on multi-turn, multi-tool interactions.
Harvard researchers challenge the standard LLM training pipeline by showing RL can be effectively applied during pre-training rather than only after SFT, finding that data composition matters more than model scale, and proposing parallel averaging of RL and SFT objectives that outperforms sequential approaches while preserving general capabilities.