Hands-on series part two: building tool use with an Agent Loop and ReAct

On November 10, 2026 at 09:29, a post titled “Hand-Building an AI Agent 02: Using Agent Loop and ReAct to let the model take its first actions” was published on cnblogs as the second installment of the “Hand-Building an AI Agent” series. The series sets out to build a general-purpose tool-using agent from scratch, with the stated goals of conversing, calling tools, planning tasks, and coordinating multiple agents, and eventually integrating into the MCP ecosystem. The stack is locked to Node.js 22 and TypeScript. This particular article concentrates on Agent Loop and ReAct, the point at which the model is first able to call tools. According to the excerpt, the model access layer uses an OpenAI-compatible interface; the excerpt cuts off before specifying which providers are compatible. The post also notes that later entries will cover task planning and multi-agent collaboration before the MCP integration. Only this single article has been collected so far, so implementation details and progress beyond this second post cannot be verified from the available sources.