Mistral Patent for “Code implemented tool calls”

A newly granted U.S. patent to French AI firm Mistral for “code implemented tool calls” is drawing criticism from developers who say it describes a basic, long‑standing pattern of LLMs generating and executing code or RPC-style tool calls. Commenters point to extensive prior art from open-source projects, academic work like CodeAct, and commercial platforms, and argue that such broad software patents mainly serve as legal weapons or bargaining chips rather than protecting genuine innovation. Some see this as another example of a broken patent system—especially in software—while others note that companies accumulate these patents defensively for cross‑licensing and deterrence, even when enforcement against big players is unlikely.

Scope and nature of the patent

  • Patent is for “code-implemented tool calls” by an LLM, filed March 2026 and already granted in the US (fast-tracked).
  • High-level claim appears to cover LLMs generating and executing code that performs tool calls, with JSON/XML-like messaging and RPC-style behavior.
  • Dependent claims emphasize a “stateless resumable sandbox” that executes generated code until non-deterministic operations (time, randomness, I/O), then replays code with cached results.

Novelty and prior art

  • Many argue this is essentially RPC/IPC or an async “await across a network,” a well-known pattern in software engineering.
  • Multiple examples of prior art are cited: Cloudflare Code Mode/MCP, Microsoft CodeAct and related paper, Anthropic/OpenAI “programmatic tool calling,” smolagents, GitHub projects, and personal systems predating the filing.
  • Some note closely related workflows where exceptions or undefined functions trigger LLM-based code generation.
  • Others concede that combining LLM self-written code with tool calls might be “novel” in a legal sense, but likely obvious to practitioners.

Critiques of software patents and the USPTO

  • Strong sentiment that most software patents, including this one, are trivial, overly broad, and mainly “patent rubbish.”
  • Several comments describe the US patent office as form- and fee-driven, leaving validity to courts; non-final rejections are acknowledged but seen as insufficient filtering.
  • Patents with prior art often still get granted and used because it’s cheaper to settle than litigate.

Motives and strategic use

  • Some see this as defensive: building a portfolio to cross-license and deter trolls or larger US players.
  • Others think it’s offensive: a potential cudgel against smaller startups and open-weight projects unlikely to afford challenges.
  • Concern that such patents might later be sold to patent trolls.

Jurisdiction, EU angle, and broader system issues

  • Repeated notes that such software patents are likely unpatentable or much harder to enforce in Europe, though loopholes via “software + generic hardware” are mentioned.
  • Example of MP3 patents is cited as precedent for European entities monetizing US software patents.
  • Broader debate on whether patents mainly serve big incumbents, hinder innovation, and should be replaced or constrained, especially for software.