Writing effective tools for LLM agents–using LLM agents (www.anthropic.com)

The Model Context Protocol (MCP) enables large language model (LLM) agents to leverage potentially hundreds of tools to solve complex, real-world tasks. This article lays out best practices for designing and refining those tools to maximize their effectiveness within agentic AI systems. Key steps include rapidly prototyping tools, rigorously evaluating them through realistic, multi-step tasks, and collaborating with agents like Claude Code for continuous iterative improvement. The process stresses validation via grounded evaluation tasks and detailed agent feedback, enhancing tool performance and usability in practical scenarios. Importantly, the piece highlights the fundamental shift in software design when working with LLM agents—tools must cater to agents’ non-deterministic, context-limited nature rather than traditional deterministic APIs intended for classic software. Effective tools must provide clear, efficient, and meaningful contextual responses optimized for token constraints while defining clear functional boundaries (namespacing). The authors advise focusing on building a few high-impact, ergonomic tools tailored to agent workflows rather than merely wrapping existing APIs, as overloaded or poorly designed tools waste an agent’s limited context window. The implications for the AI/ML community are significant: thoughtful tool design combined with agent collaboration enables scalable, intelligent agent systems that smartly integrate multiple tools to solve nuanced tasks. This approach not only advances agent capabilities but also aligns tool development more closely with how LLMs “think” and operate, pushing forward the state-of-the-art in agent-driven automation and interaction.