AI Agent Strategy Blueprint
Exactly how the XAU/USD AI Agent starter kit works — every component, every rule, and the honest limits. Signal-only, demo-only, by design.
MT4 Bridge — exports live XAU/USD data to JSON every 15 seconds
Python Engine — applies technical + news + risk filters
Telegram Alerts — signal sent to your phone, you decide the trade
By design: the agent never places orders itself. v1 is signals only.
1. Data Strategy — the MT4 Bridge
File: mt4/XAU_AI_Bridge.mq4
What it exports, and why
- 300 M15 bars — about 3 days of 15-minute candles. Enough history for EMA-50 and RSI-14 to stabilize; old enough data adds no value on M15.
- Indicators per bar: EMA-20, EMA-50, RSI-14, ATR-14 — the exact inputs the signal engine needs. Nothing more, nothing wasted.
- Refresh every 15 seconds — near-live without hammering the terminal.
- Live spread in points — feeds the spread filter (see risk rules).
- Why M15: it is the trigger timeframe of the main trading strategy (H4 bias → H1 pullback → M15 entry). The agent watches the trigger; the human confirms the higher timeframes.
2. Technical Signal Strategy
File: agent/main.py → technical_signal(). The exact rules:
| Condition | Signal | Logic |
|---|---|---|
| EMA-20 > EMA-50 and 50 ≤ RSI ≤ 68 | BUY SETUP | Uptrend + healthy momentum, not overbought |
| EMA-20 < EMA-50 and 32 ≤ RSI ≤ 50 | SELL SETUP | Downtrend + healthy momentum, not oversold |
| RSI > 68 or RSI < 32 | NO TRADE | Overbought/oversold — chasing is dangerous |
| EMA-20 ≈ EMA-50 (flat) | NO TRADE | Chop — moving averages whipsaw in ranges |
| Fewer than 60 bars | NO TRADE | Not enough data for indicators to be valid |
Honest limits
- EMA and RSI are lagging — they confirm moves, they don't predict them.
- The engine cannot see support/resistance or chart structure — that is the human's job on the H4/H1 (see the main strategy).
- In sideways markets it will flip between signals — the RSI bands (32–68) exist to reduce exactly this.
3. News Engine Strategy
File: agent/NEWS_ENGINE_SPEC.md — design for the production engine. Status in v1: placeholder (check the calendar manually until connected).
What it will ingest
CPI / Core CPIPCE / Core PCEFOMC + dot plotNFPPPIGDPISMRetail SalesJobless ClaimsFed speechesDXYUS 10Y yieldOilGold ETF flowsCentral-bank buyingGeopoliticsThe rules
- Every event graded LOW / MEDIUM / HIGH / CRITICAL.
- Blackout on HIGH/CRITICAL: no signals from 30 min before until 15 min after (matches the config).
- All events timestamped in UTC, converted to PKT for alerts.
- Store Actual vs Forecast vs Previous — the surprise moves gold, not the number alone.
- Never assume "war = gold up." Combine the headline with DXY, yields, oil, volatility and the actual price reaction.
- Multiple confirmations required before any BUY/SELL setup is produced.
4. Risk Strategy
File: agent/config.example.json — the numbers that protect capital.
Configured guards
- Risk per trade: 0.5% — position size is derived from this, never from gut feeling.
- Spread filter: 80 points max — wide spreads mean news chaos or illiquidity; signals are skipped.
- News blackout: 30 / 15 min — wired into the engine, not left to memory.
- Multi-timeframe ready: config lists M15/H1/H4/D1 — the engine will require higher-timeframe agreement before promoting a signal (planned).
- Human rule: max 2 acted signals per day; 1% daily loss cap — then the platform closes.
5. Alert Strategy
File: agent/main.py → telegram() — the loop runs every 60 seconds.
What you receive
- Signal (BUY SETUP / SELL SETUP / NO TRADE), live price, EMA-20, EMA-50, RSI, ATR.
- The reason for the signal and the news-filter state — so you can verify, not blindly follow.
- Every alert ends with: "SIGNAL/DEMO ONLY — no automatic order execution."
- Alerts never fire without a configured Telegram bot — no token, no messages, no noise.
Planned improvement: alert only when the signal changes state, instead of every loop.
6. Safety & Path to Production
Why v1 is deliberately limited — and what comes next.
Now (v1): signals only — the code physically cannot send orders to MT4.
Next: connect a real news/calendar API to the news engine.
Then: multi-timeframe confirmation (H1/H4 agreement) + paper-trade logging.
Then: 3+ months of logged demo results, reviewed weekly in the journal.
Only then: consider an execution bridge — small size, hard daily loss cap, kill switch.