AlphaAirlock: A Multi-Agent Pipeline That Vets Its Own Trading Ideas Before They Touch a Backtest

2026-07-09 · Source: Machine Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, FinTech & Digital Financial Services · Depth: Intermediate, short

Summary

AlphaAirlock is an open-source, multi-agent pipeline designed to convert unstructured market commentary into sandbox-backtested, risk-vetted trading strategies for crypto perpetuals and US equities. This system employs a five-stage LLM agent pipeline—Intake, Researcher, Coder, Backtester, and Critic—driven by an explicit state machine. Key guardrails include an AST-whitelisted sandbox for generated factor code, blocking "os", "sys", "subprocess" at the bytecode level, and an 8-second watchdog. It prevents look-ahead bias at both the factor and backtester layers, running against a bundled 2-year synthetic hourly BTC perpetual dataset (17,520 bars) with a 0.07%-per-side cost model. The Critic agent enforces hard numeric thresholds like Sharpe ratio and max drawdown, overriding LLM self-assessments. Live trading is disabled by default, ensuring a fail-closed system. The backend uses FastAPI (Python 3.11+), SQLAlchemy, and SQLite, with a Next.js frontend, and supports LLM routing via Anthropic Claude, OpenRouter, or MiniMax.

Key takeaway

For AI Engineers or MLOps Engineers building agentic systems for high-stakes domains like finance, you should prioritize robust guardrails and adversarial review. This architecture demonstrates how to prevent common failure modes, such as unreviewed code execution or look-ahead bias in simulations, by implementing explicit state machines, sandboxing, and a critical review agent with veto power. Consider adopting a fail-closed default and multi-layered validation to enhance system trustworthiness and mitigate risks.

Key insights

A multi-agent LLM pipeline can rigorously vet trading strategies through sandboxing, adversarial review, and explicit guardrails.

Principles

• Implement explicit state machines for agent workflows.
• Enforce hard numeric thresholds for critical decisions.
• Default to fail-closed for high-risk operations.

Method

The pipeline processes market commentary through Intake, Researcher, Coder, Backtester, and Critic agents, using a one-retry loop and an explicit state machine to approve or reject strategies.

In practice

• Use AST parsing to sandbox LLM-generated code.
• Apply multi-layer look-ahead bias prevention in backtesting.
• Integrate a "Critic" agent with veto power based on metrics.

Topics

• Multi-Agent Systems
• LLM Guardrails
• Algorithmic Trading
• Backtesting
• Financial Risk Management
• Code Sandboxing

Code references

• Starlight143/AlphaAirlock
• Starlight143/crucible

Best for: AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks — Fine-grained task decomposition in multi-agent LLM systems significantly improves financial trading performance.
• Agent Security Meets Regulatory Reality -- A Practitioner Systematization of Autonomous-Agent Threats and Controls in Regulated Financial Systems — Regulated financial systems amplify LLM agent security threats, demanding production-scale auditability and authorization controls.
• A Lightweight Multi-Agent Framework for Automated Concrete Barrier Design — A multi-agent framework improves LLM-based engineering design by integrating generation, evaluation, and optimization, achieving high accuracy with smaller models.
• BacktestBench: Benchmarking Large Language Models for Automated Quantitative Strategy Backtesting — Automated quantitative backtesting with LLMs requires specialized benchmarks and multi-agent systems for effective strategy evaluation.
• Building Multiple Agentic AI Trading Portfolio Pods — Autonomous agentic pods enable diversified trading, reducing emotional bias and single-strategy risk.
• Funds Coin Expands AI Trading Infrastructure Across Gold, Forex, and Stock Markets With Multi-Agent — Multi-agent AI trading systems enhance market execution by distributing strategy processing and eliminating human emotional biases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.