← All Automations AGENTIC TRADING SYSTEM

Autonomous AI Trading System

Six phases, fully autonomous — Claude Sonnet generates signals, an execution engine validates and places real paper trades on Alpaca, and a circuit breaker plus backtester make sure nothing goes sideways.

WHAT IT DOES

Signal to trade. Zero clicks.

• Claude Sonnet analyzes market data for each ticker and produces a structured Buy, Hold, or Underweight signal with a confidence score and written rationale — no hardcoded rules
• The execution engine validates every signal through five risk gates before placing a single Alpaca market-buy order: market hours, available funds, position cap, and correlation limits
• A 6-hour cron monitors every open position and places a market-sell automatically when the +10% profit target or −5% stop-loss threshold is crossed
• A circuit breaker reads portfolio equity at day-open and halts all new buys the moment it drops more than 10% — protecting against runaway drawdowns
• Every weekday evening, a digest fires via Resend email and Telegram with signal count, current P&L, and realized gains from the day

HOW IT WORKS

Six phases, one system.

1. Phase 1 — Signal Generation

   TradingAgents runs Claude Sonnet in parallel across multiple tickers. Each agent receives recent price data and produces a structured signal: Buy, Hold, or Underweight — with a 0–100 confidence score and a written rationale stored to Supabase.

2. Phase 2 — Execution Engine

   Buy signals pass through five sequential risk gates: market hours check, available funds check, per-ticker position cap, portfolio correlation threshold, and circuit breaker status. Any failed gate rejects the trade with a logged reason. Approved signals become real Alpaca paper market-buy orders.

3. Phase 3 — Exit Monitoring

   A 6-hour cron polls every open position. If current price is ≥ +10% above entry (profit target) or ≤ −5% below entry (stop-loss), the engine places a market-sell on Alpaca and writes the exit price, time, status, and realized P&L back to Supabase.

4. Phase 4 — Daily Digest

   A weekday cron fires at 9PM UTC. It reads the day's signals, current positions, and any realized P&L from Supabase, then sends a structured summary via Resend email and a Telegram bot message. One concise alert per day — no noise.

5. Phase 5 — Circuit Breaker

   At the start of each UTC day, the system reads portfolio equity and stores it as the day-open baseline in Supabase. Every time the execution engine is invoked, it checks whether current equity has fallen more than 10% below that baseline. If so, all new buys are blocked for the rest of the day — the circuit breaker fails open, so the system keeps running but stops spending.

6. Phase 6 — Execution-Logic Backtester

   An on-demand POST /backtest endpoint runs the real position-sizing and exit functions against 2 years of OHLCV data using deterministic SMA-crossover signals as a test harness. Reports return, Sharpe ratio, max drawdown, and win rate vs. buy-and-hold — validating execution logic without burning AI credits.

WATCH IT RUN

See the full system live.

Full walkthrough — signal generation, trade execution, exit logic, circuit breaker, backtester, and the live dashboard. All six phases running for real.

WHAT'S LIVE

Everything is deployed.

1. 7 API endpoints on Render

   /run · /poll-orders · /daily-digest · /breaker-status · /accuracy-report · /backtest · /health. Python 3.11.9, pinned.

2. 3 scheduled cron jobs

   Weekly signal run (Monday 9AM UTC) · 6-hourly exit polling · weekday daily digest (9PM UTC). All managed on Render's native cron scheduler.

3. 5-page Next.js dashboard on Vercel

   Overview · Signals · Positions (live Alpaca prices) · History · Performance (Chart.js P&L chart). Password-gated, reads Supabase server-side. RLS enabled — anon role denied.

4. 101 passing tests

   TDD'd across all six phases — signal generation, execution, exit logic, circuit breaker, accuracy reporting, and backtester. Run under Python 3.11 in CI.