Skip to main content

Perpetual Futures Agent

The perpetual futures agent (main.py) is the primary trading strategy. It uses a multi-agent LangGraph architecture where each trading pair runs with its own independent decision context.

Architecture Overview​

ExternalInfoAgent (Exa API market news)
         ↓
MarketData (OHLCV + technical indicators)
         ↓
EnhancedSingleSymbolAgent ←── MarketMonitor (volatility alerts)
         ↓
ExecutionAgent (structured output)
         ↓
DecisionValidator + PositionSizer + RiskManager
         ↓
ProtectionManager (plugin chain) + OrderManager
         ↓
HyperliquidClient [safety: SL fail → auto close with retries]
         ↓
SummaryAgentV2 (context compression) + ReviewAgent (learning)

Decision Pipeline​

1. Market Data Collection​

Every decision cycle collects:

  • OHLCV candlesticks at the configured timeframe (default: 1h)
  • Technical indicators: MA (20/50/200), RSI(14), MACD(12,26,9), Bollinger Bands(20,2)
  • Funding rate from Hyperliquid and Binance (CEX leading signal)
  • On-chain data: MVRV, SOPR, Fear & Greed Index (when enhanced_analysis.enabled)
  • Market news via ExternalInfoAgent (Exa API)

2. AI Decision Making​

The EnhancedSingleSymbolAgent compiles all data into a structured prompt and calls the LLM. When enabled, the decision is enriched with:

The LLM uses FinCoT 6-step reasoning:

  1. Trend confirmation — are multi-timeframe trends aligned?
  2. Entry signal — which indicators triggered? (with exact values)
  3. Sentiment check — any contrarian signals from funding/F&G/debate?
  4. Review match — which historical lesson applies?
  5. Risk calculation — stop-loss/take-profit distance, R:R ratio, fee coverage
  6. Final decision — LONG / SHORT / HOLD with confidence score

3. Execution Planning​

ExecutionAgent converts the LLM decision into a structured ExecutionPlan (Pydantic model):

class ExecutionPlan(BaseModel):
    action: Literal["LONG", "SHORT", "HOLD", "CLOSE"]
    symbol: str
    confidence: float          # 0.0 – 1.0
    leverage: int
    position_size_usd: float
    stop_loss_pct: float
    take_profit_pct: float
    reasoning: str

4. Validation Layer​

Before any order is placed, the decision passes through multiple validators:

ValidatorWhat It Checks
DecisionValidatorMulti-timeframe trend alignment, signal quality, R:R ratio, market state suitability
PositionSizerKelly criterion position sizing, volatility adjustment, drawdown shrinkage
RiskManagerATR-based dynamic SL/TP, max risk per trade (default 2%), max total exposure (50%)
ProtectionManagerPlugin chain: max drawdown (close-all + pause), daily loss (pause), consecutive loss (global or per-symbol lock), position timeout (force-close). Each plugin independently togglable via protections: config.

Validation results:

  • PASS — proceed with trade
  • WARN — log warning but allow
  • BLOCK — reject trade, log reason

5. Order Execution​

OrderManager submits orders to Hyperliquid via the SDK:

Open position → [if limit_order_enabled: limit order, else market order]
Place stop-loss → [if fails: CRITICAL SAFETY — auto close position with up to 3 retries]
Place take-profit → [if limit TP enabled]
Critical Safety Mechanism

If placing the stop-loss order fails after opening a position, HyperliquidClient immediately attempts to close the position (market order, up to 3 retries). This prevents running an open position without downside protection.

6. Context Compression​

Each trading pair maintains its own context window. SummaryAgentV2 compresses the conversation history when it grows too long, preserving key trade decisions and outcomes while reducing token costs.

7. Review & Learning​

After each closed position, the ReviewAgent records the outcome and updates the experience database. The Review System uses this to improve future decisions.

Per-Symbol Independence​

Each symbol in trading.symbols runs with:

  • Its own LangGraph state graph
  • Its own conversation history and context window
  • Its own experience/lesson database
  • Independent cooldown timers for market monitor alerts

This means BTC and ETH agents never interfere with each other's context.

Scheduling​

The agent runs on two triggers:

  1. Scheduled interval — scheduler.interval_minutes (default: 3 min) polls all symbols
  2. Volatility alert — MarketMonitor triggers an immediate decision for a specific symbol when a price spike is detected

Configuration​

Key settings in config.yaml:

trading:
  symbols: [BTC, ETH]
  max_trade_amount: 100
  max_leverage: 10

scheduler:
  interval_minutes: 3

prompt:
  set: nof1-improved

Next Steps​