[In-Depth Explanation by a Patent Attorney] Analyzing Anthropic’s “AI Agent” Patent US 12,566,913 B2 | DSL, Subtask Partitioning, 88% Confidence

💡 Key Point: This article is part of the “AI Agent Patents” series, specifically the “In-Depth Analysis of Individual Patents (Part 2)” installment. For an overview of Anthropic’s overall patent strategy, see the “Strategic Analysis” installment, and for an explanation of the sister patent, see the analysis of US 12,430,150.

Table of Contents

1. 30-Second Summary | Differences from the “Runtime” Patent
2. Basic Patent Information
3. The Core of This Patent | Four Agent Behaviors
4. Technical Highlight #1 | Custom DSL (Domain-Specific Language)
5. Technical Highlight #2 | Subtask Partitioning with Dependencies
6. Technical Highlight 3 | Feedback Loop Using Screenshots
7. Technical Highlight ④ | Quantitative Effect: “59% → 88%”
8. A Clause-by-Clause Analysis of Independent Claim 1
9. Relationship with Sister Patents | Division of Roles with US 12,430,150
10. How It Will Be Evaluated in Japanese, U.S., and European Examinations
11. Lessons for Our Own Applications | “Demonstrating Effects with Numbers”
12. Frequently Asked Questions (FAQ)

💡 Key Point: It is not the abstract idea of “having AI operate a PC,” but rather the fact that a dedicated language (DSL) was designed for this purpose that makes this patent technically concrete and grants it strong legal protection.By including the specific configuration “custom domain-specific language” in the claims, the patent avoids rejection based on an abstract idea.

💡 Key Point: This description of quantitative results is extremely important in patent practice. Specific numerical values such as “59% → 88%” significantly strengthen the argument for inventive step (particularly in Japan and Europe) compared to abstract statements like “improved accuracy.”In Japanese patent examinations, “a significant effect that a person skilled in the art could not have predicted” can be the deciding factor for inventive step, and quantitative data provides the best supporting evidence for this.

Supplementary Note | Technical Background: The parent application of this patent family refers to the multimodal architecture “Fuyu-8B” as well as training data such as recorded videos of software operations, web pages, and agentic trajectories.It is evident that this application takes into account the data infrastructure that supports the agent’s “perception and behavior.”

US 12,566,913 B2 | Claim 1 (Original Text / English)

A system for interface automation, comprising: at least one central processing unit; a memory device storing programming and data constructs that, when executed by the at least one processing unit, cause the system to configure an agent; the agent configured to: process an input including multimodal data that specifies an interface workflow, including contextual metadata and the state of the interface prior to execution of the interface workflow, wherein the multimodal data includes at least a combination of a natural language description and a prescriptive command; and segment the interface workflow into a plurality of sub-tasks, wherein each sub-task depends on the completion of a preceding sub-task; generate an output, in response to the multimodal input data, that specifies a sequence of actuation commands expressed in a custom domain-specific language (DSL) that supports both model calls and action executions, wherein the sequence of actuation commands triggers one or more machine-actuated actions that replicate user-actuated actions on the interface and automate the interface workflow by translating model instructions into real web or application events, including at least localization; actuate the sequence of multimodal actuation commands via an actuator, wherein the actuator is configured to receive the sequence of actuation commands from the agent and to perform the machine-actuated actions based on the sequence of actuation commands as synthetic actions that automate the interface workflow; and provide feedback, including interface screenshots and action histories, for iterative workflow execution and refinement.

Reference Translation by a Patent Attorney (Japanese)

💡 Key Point: It is a standard strategy to secure multiple patents for a single product (Computer Use) from different perspectives—such as the “execution platform” and the “agent itself”—to build a strong patent portfolio. By creating a structure where competitors may infringe one patent even if they avoid the other, this approach raises barriers to entry.Even for a company’s flagship product, filing multiple applications that separate the layers is an effective strategy.

We’ll assess whether your company’s AI agent can be protected by strong patents.

Patent attorneys with deep expertise in the IT, software, and AI fields provide comprehensive support—from free assessments of patentability to claim drafting based on effectiveness data, FTO searches, and filing strategies in Japan, the U.S., and Europe.

Q. What kind of patent is US 12,566,913 B2?

A. It is a U.S. registered patent held by Anthropic that protects the “AI agent” itself, which automates multimodal interfaces.The agent operates by receiving natural language instructions and screen states as input, dividing tasks into subtasks, generating a sequence of execution commands using a custom DSL (domain-specific language), and manipulating the UI. Registered on March 3, 2026, with a total of 15 claims.

Q. What is a DSL (Domain-Specific Language)?

A. It stands for Domain-Specific Language and refers to a programming language specialized for a specific domain.In this patent, the core technology is a custom DSL capable of expressing both “model calls” and “action executions” by the AI agent. The technical ingenuity lies in describing and bridging the gap between the AI’s decisions and actual UI operations using a common language.

Q. What is the difference between US 12,430,150 B1 and US 12,566,913 B2?

A. Both are sister patents that support Anthropic’s “Computer Use.”US 12,430,150 B1 protects the execution infrastructure (runtime architecture = client/server division of labor), while US 12,566,913 B2 protects the agent itself (DSL, subtask division, and feedback).This is a clever portfolio design that provides layered protection for the same product at different levels.

Q. What does the “88%” mentioned in the specification refer to?

A. The specification for this patent states that while the reliability of conventional methods was approximately 59%, the method of the present invention improved this to approximately 88%. Such quantitative results serve as strong evidence when asserting inventive step (particularly in Japan and Europe).

Q. Do you have any tips for patenting our company’s AI agent?

A. The most important lesson to be learned from this patent is to “demonstrate effects with numbers.”Rather than simply stating that “accuracy is improved,” presenting quantitative data—such as “conventional method: 59% → present method: 88%”—significantly strengthens the argument for inventive step. Additionally, it is important to specifically describe technical mechanisms, such as DSLs and subtask partitioning.

Note regarding this article: This article provides a general explanation of technology and patent systems based on published patent applications. While US 12,566,913 B2 is a registered patent, the actual scope of protection is determined by the wording of each claim, the doctrine of equivalents, and historical information.The cited claims, abstract, and specification (including reliability figures, etc.) are based on published patent bulletin data (such as FreePatentsOnline); however, for legally significant purposes (FTO, infringement analysis, invalidity, patent applications, etc.), please be sure to verify the USPTO official transcript and the latest prosecution history, and consult a specialist for an individual review.The Japanese translation is provided for reference purposes only; the official text is the original English version.