I’ve been using AI agents to build systematic strategies on Hyperliquid for over a year, and slowly built up a reliable system to represent, test and deploy strategies. Typically there are a few options 1) llm in the execution loop, ie it makes decisions and trades, but that is slow, inconsistent and hard to audit 2) you have an agent write all the code, backtest etc, but it makes tons of subtle mistakes with time series data and diff timeframe etc, and is often not reusing bits. There are some ways to improve this of course with right frameworks and setups. But its still hard to be confident in its correctness and its hard for you to verify quickly.

With Keel we built a small typed language (DSL) to represent strategies. It uses a well defined set of components to build a DAG of steps, it carries types and clocks. This allows agents to build these strategies with a really fast feedback loop as they can be quickly validated and many common errors avoided. In a DAG its also easy to inspect intermediate results. Valid strategies are fingerprinted and the exact same artifact runs in backtesting and live. These strategies are also much easier for humans to verify as they are self describing components, and mostly linear to read and reason about.

Still trying to improve the space of how many types of different strategies can be represented in this and still trying to improve how well LLMs go from natural language to strategy. A lot of the improvements come from just giving the agents better and faster tools and faster feedback loops.