8 live feeds: Binance/OKX/Bybit Perpetuals, Glassnode On-Chain, X/Twitter Stream, Telegram KOL, Coindesk, Bloomberg. ~1,240 events/second across 42 symbols. Purely technical ingest layer (WebSocket/REST).
Schema normalization, deduplication, quality tagging. Reads the full-market raw data window — the largest input consumer in the pipeline. In/out ratio ≈ 23:1.
Scans price structure across all 42 symbols. Detects breakouts, volume anomalies, regime classification (trending/ranging/volatile). Runs in parallel with On-Chain and Sentiment.
Monitors whale wallet movements, stablecoin inflows/outflows, exchange net flows. Identifies accumulation/distribution patterns across major assets.
Processes social media, KOL mentions, news narratives. Tracks hype risk and narrative momentum. Largest input in the analysis group due to high volume of social text.
Receives all 3 analysis outputs. Simultaneously constructs a PUMP thesis and a DUMP thesis — both sides mandatory. Largest output in the pipeline (85K tokens). NO SIGNAL exit point: 93% of runs stop here.
Validates evidence items — rejects low-reputation sources and unconfirmed rumors. Every rejected item saves tokens for all downstream agents.
Packages verified evidence into a standardized bundle with PUMP/DUMP scores, direction bias, and risk flags for downstream consumption.
Converts evidence into a concrete trade plan: entry range, stop loss, take-profit ladder, risk/reward ratio, leverage, timeframe, invalidation conditions. NO TRADE exit point: if no valid entry exists, the run stops here.
Queries the feature store (18-month window), finds similar historical setups. Reports win rate, profit factor, Sharpe. Statistics run on a numerical engine — LLM only summarizes.
Calculates position size based on equity, risk per trade, confidence multiplier, and volatility. Cheapest agent in the pipeline. Math runs on a numerical engine.
Reviews the trade against the BALANCED risk profile — 7 gates: size limits, leverage cap, SL enforced, daily drawdown, correlation, exposure, liquidity. Escalates to Deep Thinking if flags are raised.
The most expensive agent. Performs adversarial review — actively argues against the trade. Surfaces counter-arguments and recommends adjustments (e.g., reduce leverage). Activated only when needed: large size, high leverage, or conflicting signals.
Synthesizes trade plan + validation + risk + deep thinking → final APPROVE/REJECT. Decision is digitally signed and logged immutably. The final arbiter — can also route to manual review.
Pure deterministic code — the LLM is not allowed to improvise when real money is placed. Handles order placement, slippage control, SL/TP ladder, retries, and immutable logging. The most important design rule.
Tracks open positions, live PnL, alerts. When a position closes, lessons are written to the Knowledge Base so the Researcher Team learns. Runs periodically — tokens accumulate with holding time.
Click any run in the console and see exactly what each agent read, wrote, and cost.
Request Console Access