AgentDock Core Documentation

Core Architecture Overview

AgentDock Core provides the foundational libraries and systems for building stateful, tool-using AI agents. It emphasizes modularity, type safety, and extensibility.

Guiding Principles

  • Modularity: Key concerns (LLM interaction, state, storage, orchestration) are handled by distinct, often replaceable components.
  • Extensibility: Designed to easily incorporate new LLM providers, storage backends, custom tools, and agent logic.
  • Type Safety: Leverages TypeScript for robust typing throughout the core library.
  • Provider Agnosticism: Aims to abstract common functionalities across different LLM providers and storage systems.
  • State Management: Central focus on reliably managing conversational state across multiple interactions.

Key Subsystems & Components

AgentDock Core comprises several interacting subsystems:

  1. LLM Abstraction (/llm):

    • CoreLLM: Central class providing a unified interface to interact with various LLM providers (via Vercel AI SDK).
    • Handles streaming, function/tool calling API translation, and basic token usage reporting.

  2. Node System (/nodes):

    • AgentNode: The primary node type responsible for orchestrating agent interactions, integrating LLM calls, tool execution, and state management.
    • Defines the main processing logic for conversational agents.

  3. Storage Abstraction (/storage):

    • Provides interfaces (StorageProvider) and implementations (Memory, Redis, Vercel KV) for persistent Key-Value storage.
    • StorageFactory: Manages provider instantiation based on configuration.
    • SecureStorage: Client-side encrypted storage for browsers.
    • Foundation for planned Vector and Relational storage.

  4. Session Management (/session):

    • SessionManager: Manages session state lifecycle (creation, retrieval, update, TTL cleanup) using the Storage Abstraction layer.
    • Ensures conversational context is maintained and isolated between different sessions.

  5. Orchestration Framework (/orchestration):

    • OrchestrationStateManager: Manages session-specific orchestration state (active step, tool history, sequence progress) using SessionManager.
    • StepSequencer: Enforces defined tool execution order within steps.
    • Condition Logic: Evaluates configured conditions to control transitions between steps.
    • Controls tool availability based on the active step and sequence.

  6. Tool System (/tools, integrated in /nodes, /llm):

    • Defines tool structures and execution logic.
    • Integrates with CoreLLM for function/tool calling.
    • AgentNode handles tool loading, filtering (via Orchestration), and execution.

  7. Configuration System (/config, /templates):

    • Loads agent definitions from template files (template.json).
    • Manages environment variables for API keys and storage configuration.

  8. Error Handling (/errors):

    • Provides standardized error types and a factory for consistent error reporting.

  9. Evaluation Framework (/evaluation):

    • Provides tools for systematic evaluation of agent performance.
    • Includes EvaluationRunner, various Evaluators (RuleBased, LLMJudge, NLP, Lexical, ToolUsage), EvaluationCriteria definition, and EvaluationStorageProvider for results persistence.

Core Components

AgentDock Core consists of several key components:

  • Node System - A flexible, modular framework for building AI agents
  • LLM Providers - Unified interfaces to multiple LLM providers
  • Tool Framework - Standardized way to define and use tools with agents
  • Orchestration - Manages multi-step agent workflows
  • Response Streaming - Enhanced streaming capabilities built on Vercel AI SDK
  • LLM Orchestration Service - Bridges AgentNode and CoreLLM with orchestration capabilities

For more details on specific components:

Core Interaction Summary

The subsystems work in concert: An AgentNode uses CoreLLM for language tasks, relies on SessionManager and OrchestrationStateManager (backed by Storage) to maintain context and control flow, and executes Tools as directed by the LLM and permitted by Orchestration rules defined in the agent's Configuration.

See Architecture Overview for links to detailed documentation on each subsystem.