Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
88 lines (56 loc) · 4.1 KB

ARCHITECTURE.md

File metadata and controls

88 lines (56 loc) · 4.1 KB

CORTEX System Architecture

Technical specifications and design decisions for the CORTEX autonomous agent platform.

High-Level Architecture

The system is composed of three primary layers: the Mobile Interface, the Backend Orchestration Engine, and the Blockchain Service Layer.

  • Mobile Application: Developed with React Native and Expo, providing a dashboard for monitoring agent activity and portfolio status.
  • Backend Server: Powered by the Bun runtime and TypeScript, managing agent logic, consensus, and transaction orchestration.
  • Infrastructure: Utilizes the Groq API for AI inference and the Solana blockchain for decentralized execution.

Agent Layer

Base Agent Framework

The system utilizes an extensible agent framework. All specialized neurons inherit from a central abstract class that defines core behaviors for reasoning and participation in governance.

export abstract class Agent {
  protected groq: Groq;
  constructor(public role: AgentRole, protected apiKey: string);

  // Mandatory implementation for strategy formulation
  abstract think(context: SystemContext): Promise<Proposal>;
  
  // Mandatory implementation for democratic participation
  abstract vote(proposal: Proposal): Promise<'YES' | 'NO' | 'ABSTAIN'>;

  // Standardized AI communication interface
  protected async askGroq(prompt: string): Promise<string>;
}

Agent Specializations

Each agent is defined by a unique role, a specialized system prompt for its LLM, and a distinct voting strategy aligned with its primary objective (e.g., Yield maximization vs. Risk mitigation).

Orchestration and Consensus

Cortex Orchestrator

The orchestrator serves as the central coordination hub, executing cycles of deliberation and action.

  1. Information Gathering: The orchestrator provides agents with current market and portfolio data.
  2. Thinking Phase: Agents generate independent proposals in parallel.
  3. Voting Phase: Each proposal is subjected to a democratic vote by all ten agents.
  4. Consensus Validation: Proposals meeting the majority threshold are queued for execution.

Democratic Governance

The platform implements a majority-rule consensus model:

  • Vote Options: Affirmative, Negative, or Abstain.
  • Approval Threshold: Affirmative votes must exceed Negative votes.
  • Conservative Default: In the event of a tie, the proposal is rejected.

Blockchain Integration

Solana Service

The SolanaService class manages the connection to the Solana network, handling wallet authentication, balance inquiries on devnet, and transaction signing.

Protocol Data Layer

The PoolDataService facilitates the retrieval of real-time data from integrated DeFi protocols, including Orca, Marinade, and Kamino. This data is utilized by the agents during the thinking phase to inform their strategies.

Data Flow and Execution

  1. Cycle Trigger: The orchestrator initiates a new evaluation cycle.
  2. Strategy Formulation: Agents consult the Groq API to generate structured JSON proposals.
  3. Validation: The swarm evaluates proposals against their individual strategic goals.
  4. Execution: Approved strategies are translated into on-chain instructions via the Jupiter aggregator.

Security and Integrity

  • Environment Management: Sensitive credentials, including API keys and RPC URLs, are managed strictly through environment variables.
  • Fail-safe Logic: Agents are programmed with default "Hold" behaviors to ensure system stability in the event of API failures or ambiguous market conditions.
  • Wallet Security: Initial development utilizes disposable devnet keypairs. Production implementation will support hardware wallets and user-confirmed transaction signing.

Performance and Scalability

  • Asynchronous Execution: Agent reasoning processes are executed in parallel to minimize cycle latency.
  • Data Caching: Market data is cached with periodic invalidation to reduce RPC overhead and API costs.
  • Horizontal Scaling: The agent framework is designed to support additional specialized neurons as the platform evolves.

Last Updated: February 2026