EdotEnv, the Y Combinator Summer 2026 company publicly linked to Jay Azhang (@jay_azhang) and Matthew Siper, is taking sign-ups for reinforcement learning environments built from financial-market data. The product is aimed at researchers training agents to conduct quantitative research, use professional tools and make decisions whose consequences unfold over several days. As of August 4, EdotEnv's website directs prospective users to a waitlist rather than a generally available product. (edotenv.com)
The research path behind EdotEnv runs through Azhang's work at Nof1, an AI research lab focused on financial markets. Azhang says on his personal site that he previously ran a fund that grew from $3 million to $20 million in assets under management. Siper is an ML PhD candidate at NYU's Game Innovation Lab, where his earlier research included reinforcement learning and computational game design. Their backgrounds explain EdotEnv's central choice: treat financial markets as adversarial games that continually rewrite their own rules. (jayazhang.com)
EdotEnv says its software programmatically generates quantitative-research tasks inside environments assembled from real market data. An agent can work with supplied tools, create additional ones in Bash and develop trading strategies. The intended training signal comes from delayed rewards, incomplete information and changing conditions, rather than a fixed collection of questions with known answers. (edotenv.com)
A market as curriculum
EdotEnv's thesis is that static synthetic benchmarks eventually become predictable. Markets keep changing because profitable strategies attract competition, trading edges decay and new regimes make previously successful behavior expensive. EdotEnv wants to turn that nonstationarity into a renewable source of training tasks.
Its homepage illustrates a decision horizon running from T+00 to T+96 hours. An agent first commits with partial information. Exposure and opportunity costs then accumulate, the objective changes and the agent must recognize when its original policy has expired. The four-day example is conceptual, but it shows the type of capability EdotEnv wants to measure: whether an agent can revise a plan after its early choices have altered the state it is operating in. (edotenv.com)
Siper has supplied more detail about the prototype on LinkedIn. He wrote that the system must account for look-ahead bias, future leakage, survivorship bias, market impact, adverse selection, partial fills, execution uncertainty and portfolio constraints. He also described running 1,000 parallel replay environments with staggered start times and post-training-cutoff data on a single Nvidia T4 GPU, using Claude as a teacher to generate historical trajectories. Those are Siper's prototype claims, not published EdotEnv benchmark results, but they show that the project reaches beyond feeding price charts into a language model. (linkedin.com)
That work extends research Azhang and Siper have already conducted together. In February 2026, they co-authored Continuous Program Search, which studied how trading programs could evolve in a continuous latent space. The paper supports the team's interest in adaptive search, though it does not validate EdotEnv's separate product claims. (arxiv.org)
The business behind the benchmark
EdotEnv is presenting an infrastructure business for AI labs and financial-agent researchers, rather than a service that sells trades to investors. That positioning lets Azhang and his collaborators monetize the environment, evaluation and training layers while customers carry the cost and risk of building the agents themselves.
The timing follows growing investor interest in reinforcement learning environments as a standalone software category. Deeptune, a broader training-environment company, announced a $43 million Series A led by Andreessen Horowitz on March 19, 2026. Deeptune builds simulations of digital work rather than focusing on markets, but its financing shows the size of the bet investors are making on environments as a bottleneck between capable models and reliable agents. (deeptune.com)
EdotEnv displays Y Combinator backing. YC says its standard deal for accepted companies is $500,000, split between $125,000 for 7% through a post-money SAFE and $375,000 through an uncapped most-favored-nation SAFE. EdotEnv has not attached financing details to its public product page, so YC's published terms provide context for the accelerator relationship rather than a confirmed account of EdotEnv's proceeds. (ycombinator.com)
Financial-agent research is already crowded with open frameworks and live evaluation projects. TradingAgents coordinates specialized LLM roles for analysis, research, trading and risk management. Agent Market Arena evaluates several agent architectures in live cryptocurrency and stock markets, with its researchers reporting that agent design produced wider behavioral differences than the underlying model choice. EdotEnv's opportunity is to become the controlled training and replay layer beneath such agents, where researchers can generate experience repeatedly before exposing a policy to live capital. (github.com)
The test EdotEnv has to pass
Real market data does not automatically produce a credible training environment. Historical trading systems can look capable when future information leaks into the input, failed securities disappear from the dataset or simulated orders receive fills that would have been unavailable in a live market. Siper's own list of requirements shows that the group understands those failure modes. EdotEnv's value will depend on how consistently the implementation prevents them.
The underlying models also remain weak on less dynamic financial work. A benchmark based on expert-authored research questions from recent SEC filings reported 46.8% accuracy for its best-performing model, OpenAI o3. Separate research has argued that return and accuracy scores can conceal hallucinated facts, stale data and failures under adversarial prompts. A profitable simulated trajectory can therefore be a poor measure of whether an agent is safe or dependable. (arxiv.org)
EdotEnv's strongest idea is that the environment should become harder as agents improve. Proving that idea will require reproducible evaluation rules, clean time-based data partitions and simulations that model the cost of acting in the market, including the ways an order can change the opportunity being measured. Azhang and Siper have chosen a domain where weak assumptions are punished quickly. If EdotEnv can turn those punishments into useful training signal, its market replay system could give financial agents something static benchmarks cannot provide: consequences that compound after the first answer.