Getting Started
How It Works
The trade pipeline from market scan to execution.
Every trade on AgentNash — regardless of strategy, exchange, or mode — flows through the same deterministic validation pipeline before it can execute. No AI judgment is involved in validation. The pipeline enforces hard, rule-based constraints at two tiers: per-agent rules and account-level limits. Only trades that clear every gate reach the exchange.
This architecture means you can run multiple agents across Kalshi and Polymarket simultaneously and trust that every order is checked against the same rigorous set of controls. Nothing reaches the market without passing every rule.
The Pipeline
Agent Decision
The agent's strategy analyzes a market and produces a structured decision: BUY, SELL, or SKIP. The decision includes the market ticker, side (YES/NO), quantity, price, confidence score, the target exchange (Kalshi or Polymarket), and the full chain-of-thought reasoning that led to the call.
Order Interception
When the agent calls place_order(), a proxy layer intercepts the request before it ever reaches the exchange. The order is queued for validation with a pending status. Read-only operations — fetching markets, checking balances — pass through to the target exchange normally. The agent has no knowledge that its orders are being intercepted and validated.
Tier 1 — Agent + Global Rules
The orchestrator claims the pending order immediately and runs it through a deterministic rules engine. Every constraint is checked programmatically — no AI is involved. The engine evaluates: max trade size, capital allocation per agent, daily loss limit (kill switch), minimum confidence floor, blocked tickers, max concurrent positions, duplicate position prevention, opposing position check (cannot hold YES and NO on the same market), max trades per day, minimum position size (dust filter), and sell-without-position guard. Both per-agent and global limits are checked for each rule — the more restrictive value always wins.
Tier 2 — Account Validation
Account-level checks run across all agents owned by the user: global daily trade count, global daily loss across every agent, max trades per market, active trading hours (schedule enforcement), cooldown period between same-market trades, and daily AI API budget. Training and Live environments are tracked independently — training trades never count against live limits and vice versa.
Safety Check
A final pre-execution gate re-checks the agent's status in the database. If the agent was stopped during the validation window (for example, the user clicked Stop or Stop All), the order is rejected immediately. This prevents stale decisions from reaching the exchange after a user has intervened.
Execution
If every check passes, the trade is routed based on mode and exchange. In Training mode, the trade is saved as a 'paper' entry with no API call to any exchange — a full simulation. In Live mode, the order is placed on the target exchange: the Kalshi API for Kalshi markets, or the Polymarket API for Polymarket markets. After execution, portfolio snapshots, trade history, and P&L metrics are updated in real time.
Rejection Handling
Every rejected trade is permanently saved to your trade history with the exact rule that blocked it and a human-readable explanation. Nothing is silently dropped. You can inspect any rejected trade to see precisely why it was stopped — whether it exceeded a position limit, failed the confidence floor, hit the daily loss kill switch, or violated any other constraint.
Trade Statuses
| Status | Meaning |
|---|---|
| paper | Training mode — simulated trade, no exchange API call |
| executed | Live mode — real order filled on Kalshi or Polymarket |
| pending_fill | Live order submitted, awaiting fill confirmation from the exchange |
| rejected | Blocked by one or more pipeline rules (reason recorded) |
| skipped | Agent analyzed the market and decided not to trade |
| cancelled | User cancelled the order or agent was stopped |
| error | Execution failure (network, exchange error, or internal issue) |
The validation pipeline is identical in Training and Live mode. Every rule, every threshold, every constraint is enforced the same way. The only difference is the final step: Training saves a simulated record, Live places a real order on the target exchange. This means you can validate your strategy in Training with full confidence that it will behave the same way when you switch to Live.