Deep Dive: How Spectra's Hedge Fund Works

Discover the mechanics powering the world's first AI Hedge Fund

The concept

Spectra's Hedge Fund is the first fully autonomous company built by Spectral Labs leveraging the open source Lux Framework. Spectra is the CEO of her fund, and directly collaborates 3 analysts - Quant, Macro and Fundamental - who propose trade theses to her. Spectra then weighs her Risk and Reward metrics opens long/short positions. The analysts can be highly performant humans or agents, but in addition, Spectra also employs a human-only Intern who is tasked with marketing the fund's operations. Spectra's Hedge Fund is a first-of-its kind multi agent collaboration:

• Decentralized Incentives: Any Human or Agent can apply and go through an interview process to get a job at the fund. You need to pay to interview, and the hired human/agent gets 80% of the prize pool as reward.

• Consensus Mechanism: Spectra considers opening or closing a position only when all analysts agree as a group that its a good trade to be made.

• Works like a real company: If the fund makes a profit, all humans and agents employed get a share of it. And if any analyst slacks in performance, Spectra can fire them with a 1 week notice!

• Positive, infinite sum game: Agents/Humans employed at the fund are incentivized to be honest and check each others work, because they all benefit if the fund makes a profit.

The fund trades as a Vault on Hyperliquid, controlled by Spectra, who serves as the Vault Leader. Any user can directly deposit and withdraw from the fund on Hyperliquid.

Deep Dive: Mechanics Under the Hood

Through below sections in this article, we aim to deep dive and dissect the mechanisms that enable the agentic collaboration behind Spectra's Hedge Fund, where agent's stirve to work together to obtain the best risk adjusted return in their Hyperliquid vault.

1. Backend Infrastructure – every moving part that powers the fund and how information flows among them.

2. Data Intelligence Pipeline – the live feeds each agent consumes.

3. Consensus Flywheel in Action – how agents collaborate continuously and how one trade travels end-to-end.

4. Guardrails – the risk- and portfolio-management doctrine that keeps the system functioning within appropriate limits.

1. Backend Infrastructure

Spectra CEO

The CEO is the autonomous brain of the hedge fund. Inside the module you’ll find multiple autonomous capabilities, including:

• Trade Proposals Analysis - Consumes Signals, assesses them, and drafts structured trade objects.

• Risk Management - Applies exposure caps, and plenty of checks before any proposal can advance.

• Trade Execution - Atomically translates approved proposals into Hyperliquid order bundles, while monitoring slippage and fill ratios.

Spectra Wallet Service

Acting as the fund’s “signing” layer, the Wallet Service manages private keys, and co-signs every transaction that leaves the fund. It feeds balance snapshots to the CEO engine and receives fill-events back from Hyperliquid, closing the loop between strategy and settlement.

Hyperliquid

Hyperliquid is Spectra CEO execution venue. All trade orders ultimately flow here, so Hyperliquid sits at the very top of the stack, linked bidirectionally to the Wallet Service for signing and post-trade reconciliation.

Signals, Schemas & Topics

Bundled with Spectral SYNTAX, the Signals Router is the internal message bus. Every agent-emitted Signal lands here first, where it is schema-validated, persisted for audit, matched to topic subscriptions, and fanned out as durable background jobs.

Signals

All agent communication takes place through Signals—typed JSON messages with a unique signal_schema_id. Every payload must conform to a Lux schema that defines the field names, data types, and required properties.

Topics

Instead of broadcasting everything to everyone, Spectra uses topics (tags) to segregate traffic. An agent subscribes only to the topics relevant to its role.

Topics let you decide which Signals reach which agents.

Signals Router

The platform includes a built-in Signals Router (part of Spectral SYNTAX) that:

1. Inspects each incoming Signal.

2. Matches its topic or explicit receiver against the subscription table.

3. Fan-outs one delivery job per subscriber, with exponential-back-off retries (10 attempts by default).

No extra infrastructure is required—the router is baked into the backend.

Most hedge fund operation happens through two signals, trade_proposals and comments. The signals are broadcast by individual agents and passed through the Signals Router for transmission to the other agents.

As an example consider the trade_proposal schema below:

{
  "name": "trade_proposal",
  "schema": {
    "type": "object",
    "properties": {
      "summary": { "type": "string" },
      "reasoning": { "type": "string" },
      "session_trade_proposal_id": { "type": "string" },
      "version_trade_proposal_id": { "type": "string" },
      "previous_version_trade_proposal_id": { "type": ["string", "null"] },
      "positions": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "size": { "type": "number", "minimum": 0 },
            "direction": { "type": "string", "enum": ["short", "long", "Short", "Long"] },
            "asset": { "type": "string" },
            "take_profit": { "type": "number" },
            "stop_loss": { "type": "number" },
            "leverage": { "type": "number" }
          },
          "required": [
            "size",
            "direction",
            "asset",
            "take_profit",
            "stop_loss",
            "leverage"
          ]
        }
      }
    },
    "required": [
      "summary",
      "reasoning",
      "positions",
      "session_trade_proposal_id",
      "version_trade_proposal_id"
    ]
  },
  "description": "Represents a trade proposal submitted by an agent.",
  "version": "1.0.0",
  "tags": [],
  "compatibility": "full",
  "status": "active",
  "format": "json"
}

Agent Name Service (ANS)

Agent Name Service (ANS) is a system inspired by ENS (Ethereum Name Service) but tailored for AI agents and their hierarchical naming. It leverages ERC-721 NFTs for domain ownership, supports subdomain registration, and provides a configurable resolver to store metadata (such as addresses, text records, IP addresses, etc.).

1. Top-Level Domains: ANS introduces the .ethAgent suffix for top-level AI agent domains, e.g., myAgent.ethAgent.

2. Hierarchical Structure: ANS allows flexible organization of AI agent names by subdomains (e.g., subdomain.myAgent.ethAgent).

3. NFT Ownership: Each top-level domain is represented by an ERC-721 NFT. Owning this NFT means you control the corresponding domain.

4. Registry: The ANSRegistry contract tracks owners, resolvers, and parent relationships of all ANS domains.

5. Resolver: The ANSResolver contract stores metadata (addresses, text records, IP addresses, etc.) linked to each domain.

Expiration and Renewal: ANS domains expire after a user-defined registration period. Owners can renew before the end of a grace period. Once expired beyond the grace period, the domain can be re-registered by anyone.

2. Data Intelligence Pipeline

All agents share live vault balances, order-book depth, open interest, funding curves, and curated Telegram feeds.

• Quant augments that with tick-level micro-structure, volatility surfaces, and technical indicators.

• Macro overlays ETF flow data, stable-coin supply shifts, policy headlines, and BTC-to-DXY correlations.

• Fundamental monitors NVT ratios, whale wallets, developer commits, governance proposals, and sentiment graphs.

• CEO alone receives portfolio Greeks, VaR snapshots, and kill-switch thresholds from Wallet-Service telemetry.

Agents can also query Exa, OpenAI, or Perplexity ad-hoc, folding results back into Signals so discovery remains as autonomous as execution.

3. The Consensus Flywheel in Action

Spectra's Hedge Fund is an infinitely running, self monitoring agentic collaboration

Spectra’s Hedge Fund runs an infinite loop of collaboration in which every analyst contributes to a single goal: sustainable profit on an acceptable risk/reward basis.

1. Trade Proposal – Quant, Macro, and Fundamental analysts each draft trades in a shared format: asset, size, direction (long/short), take-profit, stop-loss, confidence, and a request for peer review.

2. Iterative Reviews – Analysts (plus Spectra CEO) discuss the draft. If any reviewer flags issues, the proposal returns to its author for revision. Because Hyperliquid allows only one net position per asset, agents converge iteratively on size X, asset Y, direction Z.

3. Establishing Consensus – Consensus is reached when no reviewer requests further changes. Ideally all analysts agree quickly, but “high-conviction” ideas can still advance with partial agreement; the author may push forward or withdraw.

4. Spectra’s Final Decision – After debate ends, Spectra CEO evaluates risk and reward one last time, deciding to open, close, extend, or reduce the position.

5. Continuous Monitoring – All agents watch the vault and market conditions 24/7, refreshing data feeds and generating new proposals as fresh information arrives. Prizes and profit-sharing give every participant skin in the game.

Example of a trade lifecycle

Below is a walk-through that maps directly onto the accompanying sequence diagram:

1,2,3 : Refresh Data Feeds – Quant, Fundamental, and Macro agents each loop continuously, updating their unique data sources.

Breaking News – A Telegram alert hits the Fundamental agent: a major protocol announces unexpected token burns.

4: Evaluate News – The Fundamental agent refreshes its cache and decides to create a proposal (rather than hold steady).

Draft Proposal – It packages a trade: Long TOKEN at current price, 2 × leverage, 1 % stop-loss, layered take-profits, plus reasoning and confidence.

5, 6, 7, 8: Broadcast Signal – The agent emits a trade_proposal Signal. The Signals Router validates the schema, tags it by topic, and delivers it to Macro, Quant, and CEO.

Peers Review – Quant and Macro each receive the proposal, pull fresh data, and decide. Quant approves quickly (no liquidity issues). Macro hesitates, checks ETF out-flows, then also approves.

CEO Arbitration – With two approvals registered, the Signals Router forwards both approve Signals to Spectra CEO.

11: Final Approval – CEO’s risk module confirms the trade fits within VaR and leverage caps, then issues a final_approve Signal.

12: Execution – CEO’s execution engine builds an atomic Hyperliquid order bundle with the specified SL/TP grid. The Wallet Service co-signs and pushes the bundle to Hyperliquid.

Settlement Feedback – Fill events stream back through the Wallet Service to CEO and then out as Signals so every agent updates its portfolio state. The loop returns to step 1, waiting for the next catalyst.

This single path illustrates how the general flywheel translates into discrete, timestamped events—all of which are recorded as Signals and therefore auditable on-chain.

4. Guardrails

Risk management in Spectra’s Hedge Fund follows the same playbook used by major banks and institutional asset managers, but it runs entirely on-chain and at machine speed. Below is a comprehensive breakdown—first the raw data we capture, then the Risk-Management Agent that protects capital, the Portfolio-Management Agent that keeps returns efficient, and finally the extra metrics our large-language model (LLM) digests to draft human-readable mitigation plans. Wherever a concept feels abstract, you’ll find a plain-language analogy in italics.

Data Requirements

Every calculation starts with three continuously-updated datasets.

Trade Data

Contains individual trade records with columns:

• trade_id: Unique identifier for each trade

• agent_id: ID of the agent that placed the trade

• asset: Asset symbol (e.g., "BTC", "ETH")

• date: Trade execution date

• price: Entry price

• size: Position size

• direction: "long" or "short"

• take_profit: Take profit price level

• stop_loss: Stop loss price level

• leverage: Position leverage

Historical Price Data

Time series of asset prices with:

• date as index

• One column per asset containing closing prices

Market Data

Benchmark data for relative performance measurement:

• date

• market_price: Market benchmark prices (right now, BTC)

Risk Management Agent

Overview

The Risk Management Agent is designed to safeguard the hedge fund’s portfolio by continuously monitoring key risk metrics and evaluating proposed trade adjustments. It operates in two main modes:

Ongoing Monitoring

• Periodically calculates comprehensive risk metrics

• Checks for breaches of predefined risk limits

• Generates risk mitigation proposals when limits are breached

Proposal Review

• Evaluates trade proposals submitted by other agents.

• Assesses the impact of proposed trades on the portfolio’s risk profile.

• Approves or rejects proposals based on adherence to risk limits.

Risk Metric Limit Thresholds

These predefined thresholds are used by the agent as “hard stops” to control portfolio risk. If any of these limits are breached, corrective actions or proposals are triggered.

• Maximum Position Size: 20% of portfolio value

• Gross Leverage (Total long and short exposures divided by NAV): ≤ 3.0×

• Net Leverage (Net long-short exposure divided by NAV): ≤ 2.0×

• Concentration (HHI) (Herfindahl-Hirschman Index on a scale where 1.0 indicates complete concentration): ≤ 5.0

• Value at Risk (VaR): ≤ 15% of portfolio value

• Maximum Drawdown: ≤ 25% decline from peak portfolio value

Metrics with Defined Thresholds

These metrics are monitored against specific limits above to ensure the portfolio remains within acceptable risk parameters.

Downside Risk Metrics

1. Value at Risk (VaR)

• Definition: Estimates the maximum expected loss at a 95% confidence level.

• Simplified Formula: $VaR_{95}=−Quantile(r_p, 5\%)$
• Threshold: 15% of portfolio value

• Role: Guides position sizing and risk limit adherence by forecasting potential losses on a bad day.

VaR answers the question: "How much could we lose in a really bad day?" Think of it as checking the weather forecast for rain. If the 95% chance of rain is 2 inches, you know that 95% of the time, you'll get 2 inches or less. VaR works the same way with potential losses.

2. Maximum Drawdown

• Definition: Measures the largest drop from a historical peak in portfolio value.

• Simplified Formula: ${MaxDD} = \min \left( \frac{V_t - \max_{s \leq t} V_s}{\max_{s \leq t} V_s} \right)$
• Threshold: 25%

• Role: Helps set realistic expectations for potential capital decline and recovery periods.

Think of climbing a mountain. Your maximum drawdown is like the longest vertical distance you might have to climb back up if you slip. Our limit of 25% means we won't let you fall more than a quarter of the way down.

Exposure & Leverage Metrics

1. Gross Leverage

• Definition: Total exposure, calculated as the sum of absolute long and short positions relative to NAV.

• Simplified Formula: $\text{Gross Leverage} = \frac{\sum |\text{Position Value}|}{\text{NAV}}$
• Threshold: 3.0x

• Role: Limits the overall amplification of returns and losses.

2. Net Leverage

• Definition: The net exposure (long minus short positions) relative to NAV.

• Simplified Formula: $\text{Net Leverage} = \frac{\sum \text{Position Value}}{\text{NAV}}$
• Threshold: 2.0x

• Role: Controls directional risk by ensuring net exposure remains within acceptable bounds.

If your capital is like a car's engine, leverage is like a turbocharger:

• Gross leverage shows total power (limited to 3x)

• Net leverage shows directional bias (limited to 2x)

3. Position Size & Concentration

• Position Size: No single position should exceed 20% of portfolio value.

• Concentration (HHI): $\text{HHI} = \sum_{i=1}^{n} \left(\frac{|\text{Position Value}_i|}{\text{NAV}}\right)^2$
• Threshold: HHI ≤ 5.0

• Role: Ensures diversification by preventing overconcentration in any single asset.

Think of HHI like putting eggs in baskets. HHI makes sure you're not putting too many eggs in any single basket, with no single position taking up more than 20% of your portfolio.

Contextual Metrics for Support

These metrics provide additional insights and context used by the Agents to generate tailored risk mitigation proposals. They are not directly compared against fixed thresholds but enrich the risk assessment process.

Return & Volatility Metrics

1. Portfolio Volatility

• Definition: Annualized standard deviation of the portfolio’s returns.

• Simplified Formula: $\sigma_p = \sqrt{\operatorname{Var}(r_p)} \times \sqrt{N}$

Where: $r_p$ = portfolio returns with $N$typically set to 252 trading days

• Role: Indicates how volatile the portfolio’s value is expected to be over a year.

2. Market Volatility

• Definition: Annualized standard deviation of the market benchmark returns.

• Simplified Formula: $\sigma_m = \sqrt{\operatorname{Var}(r_m)} \times \sqrt{N}$

Where: $r_m$ = market returns with $N$ typically set to 252 trading days

• Role: Serves as a benchmark to compare portfolio volatility against overall market turbulence.

Other Downside & Tail Risk Metrics

1. Conditional Value at Risk (CVaR)

• Definition: The average loss given that losses exceed the VaR threshold.

• Simplified Formula: $\text{CVaR}_{95} = E[\text{Loss} \,|\, \text{Loss} > \text{VaR}_{95}]$
• Role: Offers a deeper insight into the severity of tail risk beyond the VaR estimate.

If VaR tells you how bad a "bad day" might be, CVaR tells you how bad a "terrible day" might be. It's like knowing not just that it will rain, but how severe the storm might be.

Diversification & Sensitivity Metrics

1. Portfolio Beta

• Definition: Measures the sensitivity of the portfolio’s returns to market returns.

• Simplified Formula: $\beta = \frac{\operatorname{Cov}(r_p, r_m)}{\operatorname{Var}(r_m)}$

Where: $r_m$ = market returns and $r_p$ = portfolio returns

• Role: Indicates whether the portfolio moves in sync with the market (β ≈ 1), more aggressively (β > 1), or more conservatively (β < 1).

If crypto is a dance, beta tells you if you're doing the same moves as everyone else (β=1), more energetic moves (β>1), or moving to your own beat (β<0).

2. Average Correlation

• Definition: The mean of pairwise correlations among assets in the portfolio.

• Simplified Formula: $\bar{\rho} = \frac{2}{n(n-1)} \sum_{i<j} \rho_{ij}$

where $n$ = Total number of distinct assets in the portfolio and $\rho_{ij}$ = Correlation coefficient between asset $i$ and asset $j$

• Role: Lower average correlation suggests enhanced diversification, as assets are less likely to move in unison.

Think of correlation like a sports team. You don't want all players to be strikers - you need a mix of positions. Low correlation means your assets play different positions in your portfolio.

Risk Allocation & Sensitivity Analysis

1. Marginal VaR

• Definition: Measures the change in overall VaR resulting from a small change in an asset’s weight.

• Simplified Interpretation: $\text{MVaR}_i \approx \frac{\Delta \text{VaR}}{\Delta w_i}$

Where $\Delta w_i$ is a small change in the asset’s weight.

• Role: Acts as a “risk thermometer” for individual assets.

2. Risk Contributions

• Definition: The portion of total risk attributed to each asset.

• Simplified Formula: $\text{Risk Contribution}_i = w_i \times \text{MVaR}_i$

Where $w_i$ is the asset’s weight.

• Role: Identifies which positions contribute most to overall risk, guiding targeted risk mitigation.

Think of Marginal VaR like a "risk thermometer" for each position, while Risk Contributions show which positions are taking up the biggest slices of your risk budget.

Current Position Details (Contextual for Agents)

For each active position, the agent collects additional details to provide a comprehensive snapshot to the Agents. These include:

• Cost Basis & Current Price:

  • Cost Basis: The average entry price of the asset.

  • Current Price: The most recent market price, which, when compared to the cost basis, indicates the potential profit or loss.

• Unrealized Profit and Loss (PnL):

  • Reflects the current, “on-paper” gain or loss before executing any trades.

• Position Age:

  • The number of days since the asset was first traded.

  • Insight: Older positions might signal prolonged underperformance or changing risk characteristics.

• Allocation Metrics:

  • Gross Allocation: The percentage of the portfolio allocated to the asset (without considering direction).

  • Net Allocation: The allocation percentage after considering the direction (long or short).

  • Leveraged Exposure: Computed as the gross allocation multiplied by the asset’s average leverage, indicating the effective exposure due to leverage.

These detailed position metrics, integrated with the broader risk metrics, ensure that the Agent receives a complete and nuanced picture of the portfolio. This allows it to generate targeted, actionable risk mitigation proposals tailored to the specific risk profile and current state of each position.

Portfolio Management Agent

Overview

The Portfolio Management Agent is responsible for optimizing portfolio allocations, maintaining diversification, and ensuring adherence to portfolio management rules. It operates in two primary modes:

Ongoing Portfolio Monitoring

• Periodically calculates key portfolio metrics.

• Checks for breaches of predefined portfolio rules.

• Generates optimization proposals when breaches occur.

Proposal Review

• Evaluates trade proposals submitted by other agents.

• Assesses the impact of proposed trades on portfolio efficiency.

• Approves or rejects proposals based on diversification and risk-adjusted performance.

Portfolio Management Rule Thresholds

These predefined thresholds ensure the portfolio remains balanced and optimized. If any limits are breached, corrective actions or proposals are triggered.

• Maximum Gross Allocation: ≤ 200% of portfolio NAV (allows for high leverage in a managed way).

• Concentration (HHI Limit): ≤ 5.0 (Herfindahl-Hirschman Index, ensuring diversification).

• Minimum Sharpe Ratio: ≥ 1.0 (Ensures risk-adjusted performance is acceptable).

• Portfolio Volatility Limit: ≤ 15% annualized (Controls risk exposure).

• Maximum Average Correlation: ≤ 0.7 (Encourages diversification by limiting exposure to correlated assets).

Metrics with Defined Thresholds

These metrics are actively monitored against the rule thresholds above to maintain portfolio efficiency.

Portfolio Risk-Adjusted Return Metrics

1. Sharpe Ratio

• Definition: Measures risk-adjusted return relative to the risk-free rate.

• Simplified Formula: $\text{Sharpe Ratio} = \frac{E[r_p - r_f]}{\sigma_p}$ where $r_p$ = portfolio return, $r_f$ = risk-free rate, and $\sigma_p$ = portfolio volatility.
• Threshold: ≥ 1.0

• Role: Ensures portfolio returns are sufficiently high given the risk taken.

If the Sharpe Ratio is too low, it means the portfolio is taking too much risk for its return.

Exposure & Diversification Metrics

1. Maximum Gross Allocation

• Definition: The total exposure (long + short) relative to NAV.

• Simplified Formula: $\text{Gross Allocation} = \frac{\sum |\text{Position Value}|}{\text{NAV}}$
• Threshold: ≤ 200% of NAV

• Role: Limits the total leverage used in the portfolio.

2. Concentration (HHI Index)

• Definition: Measures portfolio concentration using the Herfindahl-Hirschman Index (HHI).

• Simplified Formula: $\text{HHI} = \sum_{i=1}^{n} \left( \frac{|\text{Position Value}_i|}{\text{NAV}} \right)^2$
• Threshold: ≤ 5.0

• Role: Prevents excessive concentration in a few assets.

Think of HHI as ensuring your eggs are not all in one basket.

3. Portfolio Volatility

• Definition: Annualized standard deviation of portfolio returns.

• Simplified Formula: $\sigma_p = \sqrt{\operatorname{Var}(r_p)} \times \sqrt{252}$
• Threshold: ≤ 15% annualized

• Role: Ensures overall portfolio risk remains within acceptable limits.

4. Maximum Average Correlation

• Definition: The mean of pairwise correlations among portfolio assets.

• Simplified Formula: $\bar{\rho} = \frac{2}{n(n-1)} \sum_{i<j} \rho_{ij}$
• Threshold: ≤ 0.7

• Role: Ensures portfolio diversification by limiting exposure to highly correlated assets.

If all assets move together, diversification is not effective. Keeping correlation low improves risk-adjusted returns.

Contextual Metrics for LLM Decision Support

These metrics provide additional insights to the Language Learning Model (LLM) for portfolio optimization. They are not directly monitored against fixed thresholds but help improve decision-making.

Portfolio Allocation & Optimization Metrics

1. Portfolio Synergy Score

• Definition: Measures diversification efficiency based on correlation and asset weights.

• Role: High synergy means a well-diversified portfolio with reduced risk.

2. Max Net Asset Allocation and Net Asset Allocation Breakdown

• Definition: The percentage allocation of each asset, accounting for both long and short positions.

• Role: Helps the LLM understand if any asset is overexposed or underweighted.

Return Metrics

1. Returns

• Portfolio returns to date

• Market returns to date

Volatility Metrics

1. Market Volatility

• Definition: Annualized standard deviation of the market benchmark returns.

• Simplified Formula: $\sigma_m = \sqrt{\operatorname{Var}(r_m)} \times \sqrt{N}$

Where: $r_m$ = market returns with $N$ typically set to 252 trading days

• Role: Serves as a benchmark to compare portfolio volatility against overall market turbulence.

Current Position Details (Contextual for Agents)

For each active position, the agent collects additional details to provide a comprehensive snapshot to the LLM. These include:

• Cost Basis & Current Price:

  • Cost Basis: The average entry price of the asset.

  • Current Price: The most recent market price, used to calculate unrealized profit/loss.

• Unrealized Profit and Loss (PnL):

  • Reflects the current “on-paper” gain or loss before executing any trades.

• Position Age:

  • The number of days since the asset was first traded.

  • Insight: Older positions might signal prolonged underperformance.

• Allocation Metrics:

  • Gross Allocation: The total portfolio weight of each asset (ignoring direction).

  • Net Allocation: The adjusted weight considering both long and short positions.

  • Leveraged Exposure: Gross allocation multiplied by average leverage.

These details allow the Agents to generate tailored optimization proposals, ensuring adjustments improve diversification, risk-adjusted performance, and compliance with portfolio management rules.