Context
My trading engine is written in Rust and trades real money on Hyperliquid. I am the only human working on it. Since July 2026 a fleet of Claude agents has done most of the work.
What I built
- About nine agents, each owning a single question: data quality, live-vs-backtest parity, strategy research, deploy supervision, auditing the other agents’ checks.
- Every experiment runs in its own git worktree and ends as its own PR. Only CI deploys, and only from main.
- An ops agent runs health checks, fixes what is safe and writes proposals for the rest. A deployer agent batches releases and flags stuck deploys and red gates.
- A watchdog stops the engine if the day’s loss passes a threshold.
Outcome
- The optimizer wrote about 500 hypotheses into its ledger. About 390 were killed by the gates and about 25 became merged config changes.
- One of them removed a risk filter that had never actually run in any backtest. Max drawdown of the combined book went from 43% to 27%, and Sharpe from 0.9 to 1.8 (backtest).
- Capping perp holds at 4h instead of 8h cut max drawdown from 46% to 31% at the same Sharpe (backtest).
- Rust
- Claude Code
- GitHub Actions
- PostgreSQL
- Hyperliquid API
- Telegram