Mistral Wins Patent on LLMs That Emit Pausable Code to Run Tool Calls

Mistral has been granted US Patent 12,670,045, covering a way for large language models to invoke external tools by writing code rather than emitting discrete function calls. Instead of the now-standard loop where a model spits out one JSON tool request at a time, the LLM generates a single code block that wraps several tool calls together. A server runs that block inside a sandbox and, whenever it reaches a call that must be handled elsewhere, freezes execution, ships the pending call to a client, waits for the result, splices that value back into the paused code, and resumes. The final output is handed back to the model.

The technique amounts to treating agent orchestration as ordinary program control flow, letting a model chain and compose multiple tool invocations in one pass with real branching and variable reuse instead of a chatty request-response volley. Similar ideas circulate publicly under names like CodeAct and various code-execution-with-MCP patterns, which is exactly what makes the grant notable: the approach is widely regarded as an obvious convergence point for agentic systems, and several teams have shipped comparable designs.

The significance is less technical than legal. A patent on such a broadly-applied orchestration pattern raises familiar questions about defensibility, prior art, and how aggressively it might be enforced across an ecosystem that has largely settled on tool-calling agents. For a company that leans on open-weight releases for goodwill, staking IP claims on core agent plumbing is a tension worth watching.