Language models argue, mechanical rules decide, nobody touches the keyboard.
Python
Opus/Sonnet/Haiku
Multi-Agent
Tradier
0DTE
Railway
SQLite
Discord
//FIG.02 — swarm graph
An autonomous multi-agent AI trading system. A five-agent adversarial swarm debates the market, an Opus strategist sets the morning thesis, and a 45-second execution loop trades it — while deterministic rules retain absolute veto power over every LLM.
//Role
Architect & sole engineer
//Status
Paper-trading toward go-live
//Access
Private — architecture & status only
//Problem
Discretionary trading is emotional and inconsistent; pure-LLM trading is non-deterministic and unsafe. The goal was a system that could reason like an analyst yet behave like a machine — and that no human (including its author) could nudge mid-session.
//Architecture
A five-agent adversarial swarm (Bull, Bear, Technical, Sentiment, Risk) opens each session with a structured debate. Its consensus feeds an Opus morning strategist that sets the day’s thesis. A 45-second Sonnet execution loop then operates the book. The defining choice is the LLM inversion: the models can only request trades — a deterministic layer of strategies A–F decides what actually executes.
//Role
Sole architect and engineer — agent design, the deterministic strategy layer, broker integration, persistence, and the operational guardrails.
//Safety
A circuit breaker halts new entries on daily-loss breach, ATR-based stops bound every position, and a reconciliation guard validates positions against the broker every 15 minutes to purge phantoms. The AI cannot override any of it.
//Outcome
Running unattended in a paper-trading environment, advancing toward its go-live criteria. (Private system — architecture and status only; no positions, P&L, or credentials are surfaced here.)
What it took
Technical proof.
▸5-agent adversarial swarm (Bull / Bear / Technical / Sentiment / Risk) debates before every session.
▸LLM inversion: language models only propose — deterministic strategies A–F dispose.
▸Hard safety floor the AI can never override: circuit breaker, ATR stops, position reconciliation guard.