Breaking

OSS "Auto-Quant" Lets an LLM Evolve Crypto Trading Strategies on Its Own

June 28, 2026 at 23:28 EDT

AI Agents
Open Source
Finance

An experimental open-source project called "Auto-Quant," which adapts Andrej Karpathy's "autoresearch" pattern to quantitative crypto research, has been released. An LLM agent autonomously iterates on editing, backtesting and validating trading strategies to evolve them, with staged versions and results published on GitHub.

Open Source · Agentic AI

"autotrade": an LLM that writes, tests, and discards its own crypto strategies — on a loop

A new GitHub project hands the wheel to Claude Code, OpenAI Codex, or Gemini CLI: the model reads one instruction file, then generates trading strategies, backtests them, and throws away the failures — over and over.

186

GitHub stars at release

Forks at release

100%

Codebase written in Python

LLM CLIs supported as drivers

The self-improvement loop

1 · Read

LLM reads instruction file program.md

→

2 · Generate

Writes a new strategy to generated/

→

3 · Backtest

Validates on historical data, logs to results.jsonl

→

4 · Discard

Failures moved to .trash/; loop repeats

Default experiment config & guardrails

Starting cash

100,000

Symbol · Exchange

BTC/USDT · binance

Max drawdown limit

−30%

Min profit factor

1.0

What developers like

Clear, repeatable experiment loop
Strategy file management & result logging
Works with multiple LLM CLIs, free & open
Self-correction from trade logs seen as practical

What needs caution

Needs guardrails — risk thresholds, log management
Near-autonomous flags skip permission checks
Heavy token consumption
Struggles with edge cases

The bigger shift: the LLM is no longer a mere assistant — it becomes the agent driving the iteration itself, a sign of the emerging "loop engineering" trend in agentic AI.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.