AI Agents

KEK uses a system of specialized AI agents that coordinate to analyze markets, generate strategy candidates, and cross-validate each other's reasoning.

Each agent has a distinct reasoning domain. Agents communicate through structured protocols and challenge each other's analysis — producing intelligence that no single model could generate alone.

Agents produce intelligence — not action.

What this page covers

  • The purpose of AI agents in KEK
  • Core agent roles and responsibilities
  • How agents coordinate through MCP
  • How agent outputs feed into quantitative evaluation
  • Execution and custody safety boundaries

Purpose of AI agents in KEK

KEK's agents are the reasoning core of the intelligence system:

  • Agents generate structured insights and testable strategy candidates through coordinated analysis
  • Each agent brings domain-specific reasoning (regime detection, narrative analysis, asset relevance, strategy synthesis)
  • Cross-validation between agents catches blind spots that single-model systems miss
  • Strategies must pass quantitative evaluation and paper trading before any execution is eligible

Design philosophy

KEK's agent system is built around three principles:

1) Specialization over general intelligence

Each agent is optimized for a specific domain. A regime detection agent reasons differently than a narrative agent or a strategy synthesis agent. This specialization produces deeper, more reliable analysis than a generalist model.

2) Coordination over autonomy

Agents coordinate through shared, structured interfaces. They challenge and refine each other's analysis through structured communication protocols — not by acting independently or "free-running" in production.

3) Intelligence over execution

Agent outputs are strategy intelligence — structured analysis and strategy candidates. They become eligible for execution only after quantitative evaluation and testing.

Agent architecture overview

Each agent focuses on a single domain of analysis and produces structured outputs that flow into the broader system.

Agents do not:

  • Trade
  • Allocate capital
  • Act independently in production

Their outputs are inputs to the Quant Platform for evaluation and evolution.

Core agents

Market Regime Agent

Purpose: Identify the prevailing market regime and detect transitions.

Inputs

  • Market-wide price behavior
  • Volatility and dispersion metrics
  • Structural and trend/range indicators

Outputs

  • Regime classification (e.g., trending, ranging, volatile, transitional)
  • Confidence scores and regime-shift alerts

Why it matters

This agent provides the context layer that all other agents reason against. Strategy behavior should adapt to regime — this agent tells the system which regime it's in.

Narrative Agent

Purpose: Identify macro, sector, and thematic narratives that influence asset behavior.

Inputs

  • Cross-asset structure and correlation signals
  • Liquidity and sentiment proxies
  • Contextual drivers (risk-on / risk-off behavior, thematic rotations)

Outputs

  • Narrative classifications and active themes
  • Theme relevance scoring and persistence signals

Why it matters

Markets are narrative-driven. This agent detects which stories are moving capital and how persistent they are — giving the strategy agent context that pure price data cannot provide.

Asset Relevance Agent

Purpose: Rank assets most relevant to the current regime and narrative context.

Inputs

  • Regime classification and confidence
  • Narrative theme relevance
  • Asset-level metrics (trend strength, volatility, liquidity, behavior fit)

Outputs

  • Ranked asset relevance scores
  • Context-aware filtering signals

Why it matters

This agent reduces noise and focuses strategy generation on assets most likely to express the targeted edge under current conditions.

Strategy Agent

Purpose: Synthesize intelligence from all agents into structured strategy candidates.

Inputs

  • Regime context
  • Narrative signals
  • Asset relevance scores
  • User-defined constraints and objectives

Outputs

  • Machine-readable strategy specifications
  • Parameterized rules, constraints, and assumptions
  • Variant templates for evaluation

Important

These strategies are candidates for quantitative evaluation. They must pass optimization, simulation, and paper trading before any execution is possible.

Multi-agent coordination (MCP)

Agents coordinate through the MCP (Model Context Protocol), which provides a structured system for connecting models to tools, data sources, and workflows via standardized interfaces.

MCP enables:

  • Shared context across agents and sessions
  • Structured message passing between agent roles
  • Tool coordination (data retrieval, analysis, strategy generation)
  • Versioned outputs and reproducible agent artifacts

This coordination model supports coherent multi-agent behavior without relying on hidden, mutable, or implicit state. It is fundamentally different from single-model systems where one model handles all reasoning in isolation.

How agent outputs are used

Agent intelligence flows into the Quant Platform for evaluation and evolution:

Agent Coordination → Strategy Candidates → Quantitative Evaluation → Paper Trading → Monitoring & Adaptation → Optional Execution

At no point do agents bypass evaluation or execute trades.

Boundaries & safety

KEK agents are constrained by design:

  • Do not access or custody user funds
  • Do not store or handle private keys
  • Do not self-deploy strategies
  • Do not execute trades or allocate capital

All agent outputs are treated as intelligence inputs to quantitative evaluation, monitoring, and (optionally) user-authorized execution.

Why this matters

This architecture exists to:

  • Prevent black-box behavior by keeping outputs structured and reviewable
  • Enable deeper analysis through agent specialization and cross-validation
  • Produce strategy intelligence no single model could generate alone
  • Maintain long-term trust via clear responsibility boundaries

KEK treats intelligence as input to human judgment — not as execution authority.

Where to go next