CMU AI Agents 2026: 2. Tool Use for Language Model Agents

Tool use transforms language models into agents by providing interfaces to external computer programs, enabling actions like web searching, code execution, and API interaction. These tools either extend model capabilities by accessing external data or facilitate tasks by improving speed and reliability. While programmatic tool calling via code offers high expressivity, it requires careful management of security risks and infinite loops through sandboxing. Parallel tool calling further enhances efficiency by executing independent operations simultaneously, a technique increasingly optimized through reinforcement learning. To ensure reliability, developers employ grammar-based decoding to enforce valid output schemas and utilize protocols like the Model Context Protocol (MCP) for secure, standardized integration. Evaluating these systems requires monitoring tool selection accuracy, argument validity, and overall task efficiency, as inference providers and quantization levels significantly impact performance outcomes.