I recommend checking out Anthropic's new article, "How to Build Tools for Agents." Agents differ from traditional deterministic software. Tool design must account for non-deterministic behavior and should not simply copy API or functional development approaches. It's recommended to first rapidly prototype and locally test the tool. Then, by collaborating with agents to generate a large number of real-world evaluation tasks, combine automated evaluation with agent reasoning feedback for continuous iterative optimization. Principles: Prioritize high-impact tools that cover key workflows and avoid ineffective "package API" tools. Clearly demarcate tool boundaries using namespaces (e.g., prefixes/suffixes) to reduce agent confusion. Tool responses should focus on high-signal context, prioritizing natural language and human-readable identifiers. Where necessary, support a variety of response formats, including verbose and concise, to address upstream and downstream needs. For tools that may generate large amounts of output, pagination, filtering, truncation, and optimized default parameters are recommended to avoid token waste. Error messages should be specific and clear to help agents self-correct. The tool description and parameter design should be as detailed as writing documents for new colleagues, eliminating ambiguity, strictly defining inputs and outputs, and continuously fine-tuning the description through evaluation to improve the proxy call effect.