Mistral granted US patent on pausing sandboxed code to run tool calls on the client
US patent 12,670,045 B1 issued to Mistral AI on 30 June 2026, with Gabriel Vergnaud as sole inventor, for a server that runs LLM-emitted code in a sandbox and pauses at each pending tool call to hand it to the client before resuming with the substituted result. The claim is narrower than code-emitting tool use in general — it turns on that pause-delegate-resume split, leaving agent frameworks that execute tool calls entirely server-side or entirely client-side outside its scope.
Source: patentsgazette.uspto.gov ↗
receiving, from the client, a first result of the pending tool call; resuming execution of the code block and substituting the first result of the pending tool call for the pending tool call in the code block
US Patent 12,670,045 B1, Mistral AI
Why this matters
- → Mistral's patent covers a specific architectural pattern: pausing sandboxed LLM code to offload tool execution
- → Narrows the scope of LLM tool-use patents — entirely server-side or client-side execution fall outside this cl
- → May affect multi-agent frameworks and agentic systems that split execution boundaries differently.
Pause, delegate, resume