Lightweight AI Agent Workflow System: Intelligent Orchestration Practice for Local LLMs

The open-source project ai-agent-test is a lightweight Agent workflow system built with TypeScript and AI SDK, designed specifically for local LLMs with 30B-80B parameters. It addresses the pain points of existing Agent frameworks being either too heavyweight or dependent on cloud APIs, providing complete capabilities such as multi-round iteration, tool calling, and interactive CLI, while balancing architectural simplicity and functional completeness.

With the improvement of LLM capabilities, Agent automated workflows have become mainstream, but existing frameworks have two major issues: either they are too heavyweight, or they rely on cloud APIs. A lightweight yet fully functional Agent system in the local environment has become the focus of developers. The ai-agent-test project is exactly the solution for this need.

The project's core design philosophy is "lightweight yet complete", adopting a modular architecture with core components including:

• Agent Core Loop: Multi-round iterative execution process, supporting tool calling and context management
• Tool System: Extension interfaces for file operations, command execution, etc.
• Interactive CLI: Real-time dialogue and streaming response terminal interface
• Debugging & Logging: Session logs and debugging mode The modular design allows developers to flexibly extend components without modifying the core logic.

The project focuses on local LLM support, seamlessly integrating with local inference servers like LM Studio and Ollama through endpoints compatible with the OpenAI API format, and can also connect to cloud services. The advantages include:

1. Data Privacy: All operations are done locally, and sensitive information never leaves the machine
2. Cost Reduction: Long-term usage cost is lower than cloud APIs
3. Controllable Latency: Avoids network transmission uncertainties, suitable for fast iterative workflows

Tool calling is the core capability of the Agent. The project provides a concise and powerful extension mechanism:

1. Create a tool file in the src/tools/ directory
2. Define the tool name, description, and parameter schema
3. Implement the execution logic
4. Register it in the Agent configuration This mechanism uses TypeScript static type checking to ensure extensions are intuitive and type-safe.

The project provides a fully functional interactive CLI:

• Real-time Dialogue: Automatically decides whether to call tools after inputting prompts
• Streaming Response: Model output is displayed in real time
• Debug Command: Use debug to view dialogue history
• Session Management: Use exit/quit to exit gracefully Built-in session logs automatically record interaction context, facilitating debugging and optimization.

The deployment process is concise:

1. Prepare a local LLM server (e.g., LM Studio/Ollama)
2. Install dependencies: npm install
3. Copy the configuration file: cp .env.example .env
4. Edit the .env file to configure model parameters (such as MODEL_PROVIDER, API_BASE_URL, etc.)
5. Start the Agent: npx tsx src/index.ts It takes only a few minutes from installation to running, suitable for rapid prototype verification.

Applicable scenarios include:

• Automated script development (file operations, command execution)
• Local knowledge base Q&A (combined with RAG technology)
• Development auxiliary tools (code review, document generation)
• Privacy-sensitive tasks (data does not leave the local machine) The core value of the project lies in its focus on local LLMs, modular extension, complete tool capabilities, and excellent TypeScript experience. As local LLM capabilities improve, such lightweight frameworks will play an important role in AI application development.