USPTO granted Mistral AI a patent covering LLM tool calls wrapped in sandboxed code

Sources: USPTO Official Gazette entry, discussion
Summary: US 12,670,045 B1 issued 2026-06-30, assigned to Mistral AI, filed 2026-03-04 as application 19/557,103, inventor Gabriel Vergnaud, 20 claims. Claim 1, read from the Official Gazette entry, covers receiving a request for tool calls, having an LLM generate a code block that encapsulates those calls, executing that block in a sandbox, pausing execution on a pending tool call, transmitting it to a client, then resuming and substituting the result back into the code block. That is the code-execution pattern several agent frameworks and MCP clients have adopted in place of turn-by-turn tool calling. This digest states the grant and the claim text only. It does not assess validity, claim scope, or whether any implementation infringes, and no enforcement action is known.
Why it matters: The claimed pattern is already in shipping agent tooling, so teams building on it now do so against an issued grant.