Polymarket Multi-Agent Trading

Design for one thing: positive expected value with survival (never blow up).

1) Baseline Agent Architecture

Keep each agent narrow, with strict I/O.

1. Scout (Data + News Ingest)
   • Input: market feeds, event/news feeds, market metadata
   • Output: normalized features + event triggers
2. Pricer (Fair-Value Engine)
   • Input: Scout features
   • Output: fair probability (p_fair), confidence, time-decay horizon
3. Microstructure (Orderbook Edge)
   • Input: live L2 book + trade tape
   • Output: spread state, imbalance, queue quality, fill odds
4. Arb/Consistency Agent
   • Input: related markets graph
   • Output: violations (sum probs > 1, mutual-exclusion mismatches, conditional inconsistencies)
5. Risk Sentinel (Pre-trade + Post-trade)
   • Pre: max position, max market exposure, spread/slippage gate, event blackout windows
   • Post: drawdown brakes, kill-switch, inventory skew controls
6. Executor
   • Input: approved intents
   • Output: idempotent order placement/cancel/replace, retry logic, fill reconciliation
7. Ledger
   • Output: source-of-truth positions, realized/unrealized PnL, fees, attribution
8. Atlas (Orchestrator/Judge)
   • Aggregates all outputs, scores disagreement, routes final action: buy/sell/make/hold

2) Trading Styles

Three independent alpha lanes, combined by confidence:

• A. Cross-market consistency arb (best first lane)
  • Trade when logically linked markets violate probability constraints
  • Example: mutually exclusive outcomes where aggregate implied probability is too high/low net of fees
• B. Resolution-mechanics edge
  • Trade based on rule interpretation clarity/ambiguity, not headline sentiment
• C. Maker micro-edge + rewards
  • Quote both sides where spread + expected rewards - adverse selection > 0

Avoid pure speed race as core edge; use latency only as an accelerator.

3) First Strategy: Constraint-Arb + Maker Hybrid

1. Detect clusters of linked markets (same event family)
2. Continuously compute: no-arb bounds, net edge after fees + slippage
3. Enter only when: edge > min_edge_bps, confidence > threshold, position within caps
4. Prefer maker quotes around fair value; cross spread only when dislocation half-life is short
5. Exit rules: mean reversion to fair band, rule/news invalidation, time stop, Risk Sentinel forced de-risk

4) API + Runtime Integration

• REST for snapshots/bootstrap/state reconciliation
• WebSocket for real-time book/trade/user events
• Local in-memory book per asset + periodic REST checksum
• Heartbeat watchdog; auto-reconnect with backoff and replay subscriptions
• Executor must be idempotent (client_order_id) and safe on retries
• Separate simulation, paper, and live configs

5) Hard Risk Limits (Non-Negotiable)

• Max risk per trade: 25-50 bps of equity (configurable)
• Max market exposure cap
• Max correlated-theme exposure cap
• Daily loss limit (hard stop)
• Intraday drawdown circuit breaker (soft then hard)
• No-trade mode on: stale data, websocket desync, uncertain resolution rule state, liquidity below minimum depth

6) Decision Schema

All agents must output:

{
  "market_id": "string",
  "asset_id": "string",
  "timestamp": 0,
  "action": "buy|sell|make|hold",
  "side": "yes|no|null",
  "size_usd": 0,
  "fair_prob": 0.0,
  "edge_bps": 0.0,
  "confidence": 0.0,
  "ttl_sec": 0,
  "reason_codes": ["string"],
  "risk_flags": ["string"]
}

Atlas only forwards intents that pass all risk flags. Full audit log always.

7) Backtesting & Promotion Gates

Promote strategy only if ALL pass:

1. Net positive after realistic fees/slippage
2. Positive in walk-forward (not just in-sample)
3. Robust across volatility/event regimes
4. Max drawdown within target
5. No single market family dominates PnL

8) Build Sequence

1. Data plane + book reconstruction
2. Ledger + fee-accurate PnL
3. Risk Sentinel + kill-switch
4. Constraint-Arb lane
5. Maker lane with inventory controls
6. Atlas orchestration + monitoring dashboard
7. Paper-trade 2-4 weeks before live

9) Non-Negotiables

• Never claim guaranteed profits
• Treat all strategy output as probabilistic
• Prefer explainable edges over opaque model confidence
• Log everything needed for post-trade attribution

10) Failure Modes & Recovery

11) Integration Points

• polymarket-vps-ops: Deploy and operate on VPS
• deploy-pipeline: Safe deployment workflow
• self-healing: Route tool/infra failures through retry protocol
• strategy-preflight: Stress-test before any new strategy
• trade-journal: Log all trades for pattern detection
• researcher: Resolution-mechanics edge analysis