How Quants Build Trading Agents That Improve Themselves (Full Framework)

In 2025, an AI broke a 56-year-old math record with one loop: generate, test, score, repeat. Quants run it on strategies.

Let's get straight to it....

Bookmark this

We are the team behind Horizon, the first agentic trading platform: you type a trading strategy in plain English, backtest it in minutes, and deploy it live to your broker. This article breaks down the improvement loop framework behind agentic strategy building. Currently in closed beta, launching publicly on July 15. Join the waitlist at

horizon.trade or DM

@horizon_trade_x for early access.

Here is the full framework:

• What "self-improving" actually means (and what it doesn't)
• The fitness signal: the one number the agent optimises
• Memory: fail, investigate, distil, consult
• The verifier: why the agent never grades its own work
• Where Horizon runs this loop today

What "self-improving" actually means

First, the honest version. A self-improving agent does not retrain the model. No production system does. The model's weights stay frozen.

What compounds is everything around the model: the log of tested variants, the scoring rules, the distilled lessons. The agent is a feedback loop with three parts. A generator proposes strategy variants. An evaluator scores them on historical data. A selector keeps the winners and feeds them back to the generator.

This is old machinery. Evolutionary search has been used in quant research since the 1990s. What changed is the generator: LLMs now propose variants in code and plain English, and they can run the loop for hours without a human driving each step. In 2025, DeepMind's AlphaEvolve used this exact generate-evaluate-select loop to find a faster matrix multiplication algorithm, beating a record that stood since 1969.

The fitness signal

The agent improves whatever you score. Score raw returns, and it will find the most overfit curve in your dataset. Serious desks score a composite: risk-adjusted return, drawdown, trade count, and stability across time windows.

The scoring rule is the strategy. Everything downstream is just search.

Memory: fail, investigate, distil, consult

A human researcher forgets iteration four. A well-built agent runs every failure through four stages.

It documents the failed backtest. It investigates why the variant failed: wrong regime, transaction costs, curve fitting. It distils the diagnosis into a general rule. And on the next run, it consults the rule instead of rediscovering the failure from scratch.

This is where most homemade loops break. Without this progression, the agent re-proposes variants it already rejected, and burns compute walking in circles. With it, every rejected variant marks a dead zone in the search space, and each generation starts from everything the previous ones learned.

The verifier: why the agent never grades its own work

An agent that grades its own output sees its own reasoning and prefers conclusions consistent with what it already built. In trading this failure mode has a price tag: a loop that memorizes one dataset looks like improvement on the chart and behaves like a coin flip live.

The fix has two parts. The grading rule is separate from the generator, so the process that proposed a variant never scores it. And the final grade comes from an out-of-sample gate: data the generator never saw. A variant only survives if it wins on both slices. McLean and Pontiff showed published strategies usually lose a large share of their edge once the data becomes known. Your agent's training window works the same way.

Where Horizon runs this loop

The generate, backtest, score, select loop is Horizon's core mechanic. You describe a strategy in plain English. The agent builds it, backtests it, scores it, and comes back with 2 to 3 refined variants side by side with their scores, so you pick from improved candidates instead of tweaking parameters by hand.

Every backtest feeds the next proposal. The agent does the iteration. You do the judgment.

backtest report with score

How traders get this wrong

They optimise a single metric. The agent finds the highest Sharpe in history, and it falls apart in the first live month. Composite scores exist for a reason.

They let the maker grade its own work. Ten generations of self-approved "improvement" on in-sample data is ten generations of memorisation.

They remove the human. The agent is a search engine over strategy space. It ranks candidates. Deciding what to deploy with real money stays a human call.

They confuse iteration count with progress. A thousand variants scored against a bad fitness signal is a thousand steps in the wrong direction.

Thanks for reading.

Before you go

We are the team behind Horizon, the first agentic trading platform: you type a trading strategy in plain English, backtest it in minutes, and deploy it live to your broker. The generate, backtest, score, select loop from this article is running in the product right now. Currently in closed beta, launching publicly on July 15. Join the waitlist at

horizon.trade or DM

@horizon_trade_x for early access.