Mistral has secured a patent that formalizes a method for executing tool calls generated by a large language model (LLM). The process begins when a server receives a user request to run one or more tool calls. The LLM then produces a code block in a chosen programming language, embedding the desired tool calls. The server runs this block inside a sandbox to isolate execution. When the code reaches a pending tool call, execution pauses, and the call is forwarded to the client for execution. The client returns the first result, which the server injects back into the code block, allowing the block to resume. Once finished, the server sends the final result back to the LLM, completing the cycle. This approach decouples tool execution from the LLM, enabling safer, modular interactions while preserving the model’s flexibility. It also opens avenues for integrating third‑party tools without compromising model safety, as the sandbox ensures isolation. The patent showcases Mistral’s commitment to advancing safe, extensible AI workflows.

By isolating tool calls within a sandbox, the method mitigates the risk of unintended side effects or security breaches that could arise from direct model execution. The pause-and-resume mechanism allows the system to handle asynchronous operations, such as API calls or long-running computations, without blocking the LLM. Clients can implement custom logic to validate or transform the response before it feeds back into the model’s code block. The patent further specifies how the LLM can interpret the returned results, updating the code block’s state accordingly. This design aligns with industry best practices for AI safety and could influence future standards for model-tool interactions.