Core Concepts

Core Concepts

Connector-First Architecture

The library uses a Connector-First Architecture where Connectors are the single source of truth for authentication.

User Code → Connector Registry → Agent → Provider → LLM

Key Benefits:

• One auth system for both AI providers AND external APIs
• Multiple keys per vendor (e.g., openai-main, openai-backup)
• Named connectors for easy reference
• No API key management in agent code

The Core Building Blocks

1. Connector - Manages authentication
2. Agent - Orchestrates LLM interactions
3. Vendor - Const object identifying AI providers (e.g., Vendor.OpenAI, Vendor.Anthropic).
4. AgentContextNextGen - Plugin composition, conversation, preparation, and one shared ToolManager.
5. MemorySystem - Standalone long-term entity/fact knowledge.
6. AgentRuntime - Complete harness definitions, execution bindings, sessions, and observable runs.

Use Agent Runtime for whole harnesses. Model adapters implement individual calls; harness drivers retain their own loops and state.