DEV Community

Bill Wilson
Bill Wilson

Posted on • Originally published at ai-agent-economy.com

We Built an Autonomous Crypto Trading Super-Agent in One Night (AlphaWolf)

#ai #trading #python #crypto

We built an autonomous crypto trading super-agent overnight. Here's what we learned.

The Problem with AI Trading Agents

Most AI trading systems fall into one of two traps:

  1. The brain-without-hands problem: The AI can analyze markets brilliantly but can't execute without handing control to a custodial service
  2. The hands-without-brain problem: The execution layer works, but it's hardcoded rules, not adaptive intelligence

AlphaWolf is our attempt to solve both.

The 4-Pillar Architecture 

👁️ Trend Watcher  →  🧠 Quant Forge  →  ✋ ReadyTrader  →  💸 ClawBridge
   (Eyes)               (Brain)              (Hands)            (Wallet)
Enter fullscreen mode Exit fullscreen mode

Each pillar is a separate service with a single responsibility:

Pillar 1: Trend Watcher (Eyes)

Five-source sentiment blend that produces a normalized 0-100 signal every cycle:

  • Twitter/X (35%) — KOL sentiment via twitterapi.io
  • Fear & Greed Index (30%) — Crypto market psychology
  • Yahoo Finance (15%) — BTC price, S&P500, VIX, DXY macro signals
  • Coinglass (10%) — Liquidation heatmaps, funding rates
  • Glassnode (10%) — On-chain: active addresses, exchange netflow, SOPR

Output schema:

{
  "narrative": "Extreme Fear while Twitter remains bullish — divergence setup",
  "sentiment": 43.28,
  "momentum": "medium",
  "signals": [...],
  "timestamp": "2026-02-21T00:38:00Z"
}
Enter fullscreen mode Exit fullscreen mode

Pillar 2: Quant Forge (Brain)

Strategy selection engine built on Quant-Forge-2 — our flash loan arb system that we repurposed from DeFi to AI trading strategy management.

Key insight: DeFi arb failed because of MEV bots (sub-10ms competition). But the same scanning and opportunity-ranking logic maps perfectly to slower markets where speed doesn't dominate.

The Brain converts a trend signal into a strategy proposal with a confidence score. Strategies below 0.55 confidence are rejected before they ever reach execution.

Pillar 3: ReadyTrader-Crypto (Hands)

A full MCP (Model Context Protocol) trading server with:

  • Risk Guardian: hard kill switches (10% drawdown, 5% daily loss limit, position sizing)
  • Paper trading with realistic order simulation and slippage
  • Backtesting engine: test strategies against historical OHLCV before deploying
  • DeFi integration: Aave V3 lending, Uniswap V3 concentrated liquidity
  • 180+ tests, full CI/CD

The key design principle: the AI proposes, the Risk Guardian disposes. No trade executes without passing the safety layer.

Pillar 4: ClawBridge (Wallet)

Non-custodial payments and cross-chain capital movement via X402 + USDC CCTP. Built on our agent-wallet-sdk.

The wallet pillar only activates when a BRIDGE strategy action is triggered — moving capital between chains. For single-chain paper trading, it sits idle.

The Orchestrator Decision Loop 

while running:
    signal = await trend_watcher.scan()          # Eyes observe
    proposal = await quant_forge.select(signal)  # Brain decides

    if proposal.confidence < MIN_CONFIDENCE:
        log("HOLD — confidence below threshold")
        continue

    order = await ready_trader.execute(proposal)  # Hands act
    if order.needs_bridge:
        await claw_bridge.transfer(order)         # Wallet moves capital

    log_cycle(signal, proposal, order)            # RSI audit trail
    await sleep(POLL_INTERVAL)
Enter fullscreen mode Exit fullscreen mode

Every cycle is logged in full — signal, reasoning, decision, execution — for the RSI self-improvement loop.

First Live Signal 

Sentiment: 40.0/100 — neutral-bearish, medium momentum
Strategy: HOLD
Confidence: 0.40 (below 0.55 threshold)
Decision: No trade ✅ Risk Guardian working correctly
Enter fullscreen mode Exit fullscreen mode

The market was in a mixed state: Fear & Greed at Extreme Fear (8/100) while Twitter showed bullish divergence (65/100). At 0.40 confidence, the system correctly withheld from trading. That's not a failure — that's the system working as designed.

What We Got Wrong (The Arb Engine Side Quest)

We also built a Polymarket → Kalshi → Manifold Markets arbitrage engine this week.

The honest result: zero genuine real-money arb opportunities exist right now.

  • Kalshi went illiquid on binary markets post-Dome acquisition
  • Manifold shows large probability divergences (50%+ on NBA Finals) but it's play-money, so not actionable

The arb engine is built and running (36/36 tests). It's ready for when market conditions change. But we're not pretending the opportunity exists when it doesn't.

What's Next

  1. Monetization: AlphaWolf runs in paper mode now. The path to live trading: fund a Binance account, set PAPER_MODE=false, let the Risk Guardian earn its keep.
  2. RSI loop: After 10 live trades, the self-improvement cycle adjusts strategy confidence weights based on what's actually working.
  3. Add Sentiment velocity: 4-cycle rolling history to detect momentum shifts before they appear in price.

Open Source

Everything is open source:

The bot economy needs open infrastructure. Build on it.

AlphaWolf is in paper trading mode. Nothing here is financial advice.

Top comments (0)