Anthropic recently published a detailed guide titled "Writing Effective Tools for LLM Agents—Using LLM Agents" on its official blog. The guide aims to help developers design efficient tools for large-scale language model (LLM) agents using the Model Context Protocol (MCP). The guide proposes a three-step iterative process: "Prototype-Evaluate-Collaborate" and summarizes five design principles to ensure tool effectiveness and usability.
The guide emphasizes the importance of careful tool selection, emphasizing that developers should ensure their tools precisely match the needs of LLM agents. It also recommends maintaining a clear namespace to avoid name confusion between different tools to improve development efficiency. Furthermore, the contextual information returned by tools must be more meaningful. Developers should optimize data output to increase information content and relevance, while also focusing on token efficiency to reduce transmission costs. Finally, the guide proposes improving tool descriptions through tool hints to make it easier for users to intuitively understand functionality and usage.
Notably, many of the conclusions in the guide are automatically generated by Claude Code after analyzing scripts and refactoring tool descriptions, ensuring scientific accuracy. To prevent overfitting, Anthropic maintains a test set for evaluation and has open-sourced the tool evaluation cookbook. Officials also stated that with future upgrades to the MCP protocol and underlying LLM, tool capabilities will continue to evolve, providing stronger support for developers.
