EdotEnv, a startup from the Y Combinator S26 batch, was founded by two former quantitative traders. The company is building training environments for AI agents using real financial market data. It has open-sourced a sample task repository on GitHub and plans to sell continuously improving environments to AI labs, researchers, and companies that train their own AI agents. (edotenv.com; Hacker News)

The idea is straightforward. Most tests for AI involve fixed problems — a set of math questions, or coding challenges. Once an AI gets good enough at those, the test stops being useful. Markets are different. They generate new data every day, they are messy and unpredictable, and they give immediate feedback: a strategy either makes money or loses it. No human expert is needed to grade the result. (edotenv.com; Hacker News)

In EdotEnv's environments, AI agents get professional tools and can build their own tools using a command-line interface. Tasks include building predictive models, designing investment portfolios, testing strategies against historical data, and adapting to changing market conditions. Each task comes with a testing tool, an execution tool, and a final evaluation step. The scoring system is designed to measure how well the agent builds useful features while still taking advantage of market properties. (Hacker News)

The founders, Rui and Michael, come from quantitative trading backgrounds. Their Y Combinator affiliation places EdotEnv in the S26 batch. (Hacker News)

EdotEnv's own findings from running today's best models through these environments are notable. The models struggle to iterate deeply on research ideas. Giving them more reasoning effort does not improve performance. And when agents start losing money, they stop trading rather than adjusting their approach. (Hacker News)

These findings challenge a common assumption in AI research right now. The prevailing belief is that if you give an AI model more reasoning steps and a bigger thinking budget, it will do better on complex tasks. In a market environment full of noise and long feedback loops, that does not appear to hold. The tendency of agents to disengage under losses rather than adapt is a specific failure mode that fixed tests would never catch, because they rarely punish inaction in a meaningful way.

The sample repository, available at github.com/MMcollab-dotcom/feature-engineering, lets researchers inspect the task structure, tooling, and reward design directly. (Hacker News)

The business model targets a real gap. AI labs training agents need environments that do not get solved and then become useless. Most existing tests, whether for coding, math, or general reasoning, eventually see top models reach or exceed human-level performance, at which point the test stops providing useful signal. Markets offer a domain where the problem space regenerates itself. EdotEnv's plan to sell continuously improving environments positions it as an infrastructure provider for agent evaluation. (Hacker News)

Using financial markets as a training ground carries real limitations. Market data is noisy, shifts over time, and can enter prolonged periods where old patterns stop working. An agent trained to perform well under one set of conditions may fail badly under another. This is a known problem in quantitative finance, and EdotEnv's scoring system, which tries to separate the agent's skill from raw market exposure, is an attempt to manage it. Whether that separation holds across diverse market conditions remains an open question.

There is also a broader question about whether skills learned here carry over to other fields. EdotEnv frames its environments as teaching applied machine learning and long-term planning under messy, adversarial conditions. Those are general skills. A model that learns to build predictions, test them, and adapt to changes in a financial context may transfer those abilities to other domains. It may also overfit to the specific patterns of market data. The company's own findings, showing that current models struggle to iterate deeply on research ideas, suggest that any such transfer is not yet happening at a level the founders find compelling.

The competitive landscape is active. Several efforts, both academic and commercial, are working on new ways to test AI agents that go beyond fixed problem sets. EdotEnv's differentiation is its use of real market data, which provides natural rewards and avoids the ceiling that hand-designed tests hit. The trade-off is that real market data introduces confounding factors that synthetic environments can control more precisely.

For AI labs, the appeal of a test that never saturates is direct. Training systems that rely on benchmark progress need fresh challenges as models improve. A market-based environment that regenerates its own difficulty could extend the useful life of an evaluation pipeline. For quantitative finance firms, the appeal is different: an agent that can build features, test them, and adapt on its own is essentially a compressed version of a quant research workflow.

EdotEnv is early. The sample repository is public, the full environment set is not yet commercially available, and the founders' own results show current models falling short of the deep iteration the tasks demand. What is on offer is a clear methodology, a concrete open-source artifact, and a business thesis that addresses a genuine problem in how we evaluate AI agents.