quant-agents: Autonomous Quant Research on AMD

quant-agents is a local "Quant Hedge Fund in a Box": a multi-agent system that runs the full quant research loop end to end with zero cloud dependency. A LangGraph state machine routes work between four agents. The Ideator proposes a market hypothesis with an economic rationale, precise statistical entry and exit rules, and regime filters. The Coder turns it into a fully vectorized pandas/numpy generate_signals function (no loops, no iterrows). A sandbox executes the code on SPY daily data, lags every signal with .shift(1), and computes Sharpe, max drawdown, total return, and win rate. The Auditor acts as head of risk: it diagnoses tracebacks, rejects lookahead bias and curve-fitting, and demands structural fixes like regime filters or volatility-based sizing instead of parameter tweaking. A strategy only passes with Sharpe above 1.5 and drawdown under 15%, checked in code, not by the LLM. Survivors go to the Packager, which writes a production handoff report with documented code, Big-O analysis, execution edge cases, and a C++ translation plan (Eigen/Armadillo). The whole system runs on a single Qwen3.6-35B-A3B Mixture-of-Experts model served by vLLM on AMD Instinct MI300X with ROCm. Instead of loading four models and fragmenting HBM, all agents share one model and LangGraph swaps the persona per call. With only ~3B active parameters per token, the agent loop stays fast enough to iterate many times per strategy. Circuit breakers cap rework attempts and new ideas, so the system either finds a strategy that survives the audit or stops cleanly.