Agentic: A Modular Component Library for Building AI Agent Workflows

Agentic is a TypeScript modular component library for AI agent workflow development, designed to solve core issues in agent development such as tool management, context tracking, and state maintenance. It adopts an atomic component design, follows the Unix philosophy (each component does one thing well), supports multiple model providers, emphasizes type safety and observability, and provides reliable infrastructure for developers.

AI agent development faces challenges like tool management and context handling. Agentic is not a complete framework but an atomic component focused on specific problems, allowing developers to combine components as needed and avoid the complexity of large frameworks. Its core focuses on three closely related areas: tool registration and management, context injection and deduplication, and conversation state tracking.

ToolRegistry addresses tool availability and safe execution issues. A tool definition includes name, description, parameters, category (e.g., file operations), and cost level (cheap/expensive), helping agents balance resource consumption. The execute method encapsulates call logic and has built-in usage statistics (count, failure rate), supporting export of tool definitions to OpenAI/Anthropic formats for seamless multi-model switching.

ContextManager solves information management issues in limited context windows. Each context item carries metadata such as content, ID, category, priority, and timestamp, supporting complex query filtering. It provides three deduplication mechanisms: exact ID matching, content hash detection, and fuzzy similarity recognition to prevent duplicate information pollution. The position parameter supports conversation position awareness, enabling intelligent context compression (e.g., summarization of old information).

Agentic leverages the TypeScript type system to enhance the development experience. The Tool interface balances flexibility and safety (parameters use JSON Schema style). The execute method supports asynchronous operations and context transfer. The library design considers testing and monitoring needs, providing features like real-time usage statistics, batch execution, and category filtering to help diagnose issues and optimize performance.

Agentic uses a modular architecture. Execution sub-packages (e.g., execution-openai) independently encapsulate model adapters, making it easy for the community to extend support for other models. It does not restrict conversation flows, planning strategies, or tool implementations. As an underlying infrastructure, it can work with various agent systems, providing a solid foundation for custom agents.

Agentic is suitable for production systems that require fine-grained tool control, multi-model switching, and high observability. Novices are advised to first build a clear tool set (ToolRegistry) before introducing context management. Existing systems can replace components incrementally (e.g., replace the tool system first) without full reconstruction.

Agentic represents the maturity of the agent toolchain, focusing on solving basic problems rather than being an all-in-one framework, reflecting the progress of the field from exploration to engineering practice. As agents move from prototypes to production, such infrastructure will become more important. Agentic provides well-thought-out solutions to basic problems like tool management and context control, and is worth developers' attention and trial.