AgentNash
What is AgentNash?How It WorksConnecting Your AccountTraining vs Live ModeStrategiesCouncil V2The SuperforecasterThe Council (Legacy)TerminalAgentsBenchmarkingSafeguards & RulesNuclear Option

Strategies

The Council (Legacy)

The original 5-agent parallel ensemble for Kalshi and Polymarket.

The Council (V1) is deprecated. The Council V2 is the active replacement with improved architecture, tighter risk controls, and lower latency. This page is preserved as a reference for users reviewing historical V1 trades.

The Council V1 is a 5-agent parallel ensemble strategy that operated on both Kalshi and Polymarket. Five specialized AI models — each from a different provider — independently analyze every market opportunity. Their probability estimates are combined through a confidence-adjusted weighted average to produce a consensus probability, which is then compared against the market price to identify tradeable edges.

Unlike single-model approaches, the Council forces diversity of perspective. A forecaster anchors the base rate, a news analyst scores recent sentiment, a bull makes the strongest YES case, a bear counters it, and a risk manager evaluates expected value. The weighted consensus reduces hallucination risk, overconfidence bias, and single-point-of-failure errors.

Pipeline at a Glance

Every market opportunity flows through 8 stages before any capital moves:

1

Market Scanning

Fetch active binary markets from the exchange API. Filter by volume, category, and status. Repeat every 30 seconds.

2

Selection & Dedup

Run 6 pre-checks to avoid wasting AI credits on markets already analyzed, below budget, or on cooldown.

3

Ensemble Analysis

5 AI agents analyze the market in parallel: Forecaster, News Analyst, Bull Researcher, Bear Researcher, and Risk Manager. Their outputs are aggregated into a weighted consensus.

4

Edge Calculation

Compare the consensus probability against the market price. Required edge depends on confidence: 6% at high confidence, 8% at medium, 12% at low. Below 50% confidence always skips.

5

Position Sizing

A tier-based system determines how many contracts to buy based on account balance, confidence edge, and Kelly Criterion.

6

Risk Checks

Position limits, cash reserves, and the backend rules engine validate the trade. Any failure blocks execution. See Safeguards for details.

7

Order Execution

The order is intercepted, queued, and processed by the orchestrator. Training mode saves a paper trade. Live mode places a real order on the exchange.

8

Settlement & Exit

Open positions are monitored with dynamic stop-loss and take-profit levels. Auto-exit after max hold time. Settlement checker polls for resolved markets.

The 5 Agents

All five agents run in parallel. Each agent produces an independent probability estimate and confidence score. The ensemble aggregates these into a single weighted-average probability. A minimum of 3 successful agent results is required to produce a consensus.

RoleModelWeightPurpose
ForecasterGrok 4.1 Fast (xAI)30%Estimates true YES probability using base rates and historical patterns
News AnalystClaude Sonnet 4.6 (Anthropic)20%Scores recent news sentiment and relevance to the market
Bull Researchero4-mini (OpenAI)20%Makes the strongest evidence-based YES case with 3-5 arguments
Bear ResearcherGemini 3.1 Pro Preview (Google)15%Counters the Bull with counter-evidence and risk factors
Risk ManagerDeepSeek V3.2 (DeepSeek)15%Calculates expected value, risk score, and recommended position size

Weights and Consensus Logic

The ensemble computes a confidence-adjusted weighted average. Each agent's base weight is multiplied by its self-reported confidence (floored at 0.1 to prevent zero-weight), so agents that are more confident about their estimate contribute more to the final probability.

Formula: weighted_prob = sum(prob_i * weight_i * max(conf_i, 0.1)) / sum(weight_i * max(conf_i, 0.1))

If the standard deviation of agent probabilities exceeds 0.25 (the disagreement threshold), a confidence penalty of up to 30% is applied. High disagreement signals uncertainty, and the system responds by reducing conviction rather than forcing a trade.

Agent Role Details

Forecaster (Grok 4.1 Fast)

Estimates the true probability that the market resolves YES using a structured 5-step method: establish the base rate, assess current conditions, check market structure (single event vs. parlay), calibrate toward the base rate when uncertain, and flag edge only when the estimate differs from market price by more than 10 percentage points. Must not fabricate statistics.

Output: probability, confidence, base_rate, side, reasoning.

News Analyst (Claude Sonnet 4.6)

Scores recent news sentiment from -1.0 (strongly bearish) to +1.0 (strongly bullish) and relevance from 0.0 to 1.0. Only the last 24-48 hours of developments are treated as meaningful. The News Analyst's probability is derived from sentiment and relevance: prob = 0.5 + (sentiment * relevance * 0.5).

News sources: Category-aware RSS feeds (NYT, BBC) plus optional Perplexity Sonar Pro for real-time web search.

Bull Researcher (o4-mini)

Makes the strongest evidence-based YES case with 3-5 concrete arguments. Estimates a probability_floor — the minimum reasonable YES probability even if the bear is right about some things. Identifies near-term catalysts that could push probability higher.

Bear Researcher (Gemini 3.1 Pro Preview)

Counters the Bull's specific arguments with 3-5 concrete reasons the event is unlikely. Estimates a probability_ceiling — the maximum reasonable YES probability. Highlights risk factors and historical precedents where similar events failed.

Risk Manager (DeepSeek V3.2)

The quantitative backbone. Calculates expected value, assigns a risk score (1-10), recommends position size using fractional Kelly criterion, and evaluates edge durability.

  • EV for buying YES at X¢: (your_probability * $1.00) - $0.X
  • EV for buying NO at Y¢: ((1 - your_probability) * $1.00) - $0.Y
  • Kelly sizing: size = (edge / odds) * kelly_fraction

Edge Calculation

After the ensemble produces a consensus probability, the agent calculates the edge — the difference between the AI's estimate and the market price. The edge must be in the same direction as the proposed trade.

Consensus ConfidenceEdge RequiredReasoning
80%+6%+High certainty — smaller edge is still worth taking
60-79%8%+Medium certainty — need a clearer mispricing
50-59%12%+Low certainty — only trade if the edge is obvious
Below 50%Always skipConfidence too low regardless of edge

Position Sizing

A tier-based system adapts position size to account balance. Smaller accounts allocate a larger percentage per trade; larger accounts allocate less to protect capital.

Account SizeBase %Max %Max Contracts/Order
Under $10020%40%10
$100 - $1K5%15%50
$1K - $10K3%8%250
$10K - $100K2%5%1,000
$100K+1%3%5,000

The base investment is scaled by the confidence edge: investment * (1 + multiplier * edge), capped at the tier maximum. After the tier calculation, the Risk Manager's recommended size acts as an additional cap. Quarter-Kelly (kelly_fraction = 0.25) is used in production for safety.

Risk Checks

Every trade must pass three layers of deterministic validation: agent-level position limits, the backend 11-rule engine, and account-level caps. Any single failure blocks execution. All thresholds are configurable from Settings → Safeguards. See the Safeguards page for the full rule-by-rule breakdown.

Execution & Settlement

The agent never talks directly to the exchange. Every order is intercepted, queued for validation, and processed by the orchestrator. Training mode saves a paper trade; live mode places a real order. SKIP decisions are also recorded for a complete audit trail.

Open positions are monitored with dynamic stop-loss and take-profit levels. Positions auto-exit after max hold time (default: 240 hours). A settlement checker polls the exchange for resolved markets and updates P&L.

Polymarket Differences

The same 5-agent ensemble runs on Polymarket with identical AI logic. The differences are in how markets are fetched, how orders are signed, and how settlements resolve. The backend detects the target exchange from the agent configuration and routes accordingly.

Market Data

The Polymarket variant fetches markets from the exchange data API and order book data from the exchange API. Markets are identified by condition IDs and token IDs rather than Kalshi tickers.

Order Signing

Polymarket orders are signed using cryptographic wallet signatures on the Polygon network. Credentials are handled securely as needed from the user's private key. Wallet-based signing is used for order authentication.

Settlement

Polymarket uses the UMA Optimistic Oracle for outcome resolution, which introduces a dispute window before finality. The settlement checker accounts for this delay when polling for resolved markets.

Kalshi vs. Polymarket Comparison

AspectKalshiPolymarket
AuthCryptographic signature authenticationWallet-based signing
ChainOff-chain (centralized)Polygon
SettlementKalshi internal resolutionUMA Optimistic Oracle
Price formatCents (1-99)Decimal (0.01-0.99)
Market IDsTicker stringsCondition ID + Token ID

Models & Costs

All AI requests route through OpenRouter. The ensemble model roster as configured in V1:

ModelProviderRoleWeight
Grok 4.1 FastxAI (via OpenRouter)Forecaster30%
Claude Sonnet 4.6Anthropic (via OpenRouter)News Analyst20%
o4-miniOpenAI (via OpenRouter)Bull Researcher20%
Gemini 3.1 Pro PreviewGoogle (via OpenRouter)Bear Researcher15%
DeepSeek V3.2DeepSeek (via OpenRouter)Risk Manager15%

A typical full ensemble pass costs approximately $0.02 - $0.08 depending on market complexity and response lengths. The system includes automatic model failover if any provider is temporarily unavailable.

All Configurable Settings

Agent-level settings are in the agent's configuration. Account-level settings are in Settings → Safeguards.

Market Scanning

SettingDefaultDescription
Min VolumeTiered (50-1,000)Minimum contracts traded, scales with account balance
Scan Interval30sSeconds between market scan cycles
Max Expiry30 daysMax days to expiry for eligible markets
Preferred CategoriesAllWhitelist (empty = trade all)
Excluded CategoriesNoneBlacklist (empty = exclude none)

Ensemble & AI

SettingDefaultDescription
Min Models for Consensus3Minimum successful agents to produce a consensus
Disagreement Threshold0.25Std dev above this triggers confidence penalty
Parallel RequeststrueRun agents in parallel (vs. sequential fallback)
Max Ensemble Cost$0.50Max cost per ensemble decision
Daily AI Budget$10Daily spending limit on AI API calls
Analysis Cooldown3 hoursMin hours between same-market analyses
Max Analyses per Market/Day4Max analyses per market per day
AI Temperature0Model temperature (0 = deterministic)
AI Max Tokens8,000Max tokens per model response

Position Sizing & Risk

SettingDefaultDescription
Tier systemAutoBase/max percentages adapt to account size
Position Size Multiplier1.0Multiplier for edge-based scaling
Kelly Fraction0.25Quarter-Kelly for production safety
Max Single Position30%Max % of portfolio per position
Max Positions3Max concurrent open positions
Min Balance$5Minimum balance to start trading

Trading Frequency

SettingDefaultDescription
Max Trades per Hour20Rate limit on trade execution
Run Interval10 minHow often the trading cycle runs
Position Check Interval15sHow often open positions are checked
Processor Workers5Concurrent market analysis threads

Exit & Risk Management

SettingDefaultDescription
Profit Threshold20%Take profits at this return
Loss Threshold15%Cut losses at this drawdown
Max Hold Time240 hoursAuto-exit after 10 days
Confidence Decay Threshold25%Exit if confidence drops this much
Max Daily Loss %20%Daily loss kill switch (% of capital)

Prediction market trading involves real financial risk. Past performance does not guarantee future results. The Council V1 is deprecated and no longer receives updates. If you are starting fresh, use The Council V2 instead.