In this article, a patent attorney specializing in AI intellectual property will provide an in-depth analysis of this patent—including its full claim structure, technical architecture, and the ingenuity of the specification—while citing the original text. This is a must-read for developers and IP professionals who are developing their own “Computer Use”-type agents and want to know “what the key points for patent protection are.”
💡 Key Point: This article is part of the “In-Depth Analysis of Individual Patents” series within our AI agent patent series. For an analysis of Anthropic’s overall patent strategy, please see “Decoding Anthropic’s Patent Strategy,” and for the basic requirements, please refer to the “Basics” section.
Table of Contents
| Item | Content |
|---|---|
| Patent Number | US 12,430,150 B1 |
| Title of the Invention | Runtime architecture for interfacing with agents to automate multimodal interface workflows |
| Registration Date | September 30, 2025 |
| Filing Date | October 8, 2024 |
| Priority Date | March 20, 2024 |
| Applicant | Anthropic PBC |
| Inventors | Rohan Bavishi, Erich Elsen, Curtis Hawthorne, et al. |
| Number of Claims | 20 (3 independent claims: Claims 1, 14, and 20; 17 dependent claims) |
| Status | Granted Patent |
“Computer Use” is a technology that enables AI to operate any software by viewing the screen (recognizing screenshots), moving the mouse, clicking, and entering text—just as a human would. Unlike traditional automation via APIs, this technology operates directly on the UI itself, making it groundbreaking because it can automate even apps that do not expose their APIs.
This patent protects the underlying architecture that enables this “AI-driven UI automation” at runtime. Please note that it secures the rights to the mechanism that makes this work, rather than just the idea itself.
The most distinctive feature of this patent is its architecture, which divides processing between the client and server sides and connects them via an abstraction layer called an “intermediate representation.” The overall flow is as follows.
💡 Key point: By inserting an abstraction layer called “intermediate representation,” the agent’s specifications (what it wants to do) are separated from the actual UI operation commands (how to operate it).This approach is similar to compiler design, elevating the simple idea of “having an AI operate a PC” into a concrete software architecture.
According to the specification, the “agent functions” that an agent can call are systematized into several types. These serve as a specific operational vocabulary that underpins the invention.
| Type | Examples of Functions (as described in the specification) | Role |
|---|---|---|
| Built-in Functions | answerQuestionAboutScreen (answering questions about the screen), click, type, scroll | Basic UI Operations and Recognition |
| Planner Functions | act (action), fillform (form input), pickdate (date selection) | Higher-level tasks combining multiple operations |
| Workflow Features | (User-defined Workflows) | Business-Specific Automation Units |
This architecture, in which operations are abstracted as functions and AI agents combine them to perform tasks, is also specifically outlined in the dependent claims (discussed later).
In order for an AI agent to determine “what to do next,” it must perceive the current situation. The specification of this patent states that the observation logic provides the following to the agent:
| Observation Information | Content |
|---|---|
| Screenshot | Visual information from the current screen (multimodal input) |
| Action History | History of operations performed to date |
| Task description | Description of the task to be accomplished |
The perception-action loop—“view the screen → consider past actions → decide on the next action”—is the very basis for an AI agent’s “autonomy.” This patent technically implements this loop.
This patent consists of a total of 20 claims and follows the standard structure for software patents, with independent claims organized into three categories: methods, systems, and media.
| Claims | Category | Subject Matter | Presumed Infringer |
|---|---|---|---|
| Claim 1 | System | A system operating on a processor | A person who manufactures, uses, or sells the device |
| Claim 14 | Computer-implemented method | Processing procedure | A person implementing said method |
| Claim 20 | Non-transitory computer-readable storage medium | A medium on which a program is recorded | A person who distributes or provides the program |
The dependent claims (2–13, 15–19) specifically define the type of agent function, observation logic, return values, and other elements, creating a multi-layered defense to ensure that rights remain even if the independent claims are deemed invalid.
Claim 1 (Original Text / English)
A system, running on one or more processors, for client-side implementation of an interface automation language at runtime, comprising: agent specification logic, running on the client side and configured to construct an agent specification and to make the agent specification available for server-side translation into an intermediate representation, wherein the agent specification is configured to automate a multimodal interface workflow; and runtime interpretation logic, running on the client side and configured to: receive the intermediate representation; detect one or more agent functions in the intermediate representation; generate one or more agent calls based on the agent functions; issue the agent calls to an agent, and, in response, receive at least one runtime actuation function from the agent; and translate the runtime actuation function into at least one runtime actuation command, wherein the runtime actuation command triggers at least one machine-actuated action as a runtime synthetic action that automates the multimodal interface workflow.
Reference Translation by a Patent Attorney (Japanese)
| Limitations | Technical Meaning | Reason for Effectiveness |
|---|---|---|
| Translation to an intermediate representation | An abstraction layer that separates specification from execution | Moving Beyond Abstract Ideas: Architectural Concreteness |
| Detection and Invocation of Agent Functions | Functionalization and Dynamic Invocation of Operations | Concretization of control logic. Key points for demonstrating progress |
| Translation from operational functions to operational commands | Converting AI Decisions into Actual UI Operations | The Core of the Technical Implementation of “Screen Manipulation” |
Claim 14, a method claim, describes a process nearly identical to that of Claim 1 in the form of a method, but there is one important difference.
Claim 14 (Excerpt) (Original Text / English)
... issuing, on the client-side, the agent calls to an agent on the server-side, and, in response, receiving, on the client-side, at least one runtime actuation function from the agent; ...
While Claim 1 (system) simply states “issuing the agent calls to an agent,” Claim 14 (method) and Claim 20 (medium) explicitly state “to an agent on the server-side.”
💡 Key Point: This difference is a deliberate choice in claim drafting. By limiting the scope of protection to client-side devices only, the system claim broadly captures “those who implement the client-side.”On the other hand, the method and medium claims explicitly specify coordination with a “server-side agent,” thereby covering the processing of the entire distributed system. This is a prime example of providing layered protection for the same invention from different angles.
① It corresponds one-to-one with the product. It directly protects the actual core function—“Computer Use”—and its business value is clear.
② It protects the architecture, not just an idea. Rather than an abstract concept like “AI operating a PC,” it is described as a concrete software structure involving intermediate representations, functionalization, and multi-stage transformations, resulting in high patent eligibility.
③ The claims are structured in a multi-layered manner. In addition to the three categories—system (client-side focused), method, and medium—the claims vary the description of the agent’s location (server-side) between system and method claims, thereby capturing different forms of infringement.
The Alice/Mayo tests examine whether an invention consists of an “abstract idea” or involves an “inventive concept.” Since this case involves technical implementations such as intermediate representations, agent invocation, and the translation of operation commands, it is easier to argue that it provides a “technical solution to a technical problem (automated UI operation),” and the patent is currently registered.
The criterion is whether “information processing is concretely implemented using hardware resources.” Since the specification describes a system running on a processor, client/server division of labor, and specific data conversion, the configuration is well-suited to meeting patent eligibility requirements as a software-related invention.The key to demonstrating inventive step lies in the technical ingenuity of “abstraction through intermediate representations” and “conversion of operational functions to commands.”
Technical contribution (COMVIK) is assessed. This configuration is easily positioned as an architectural technical solution to the technical challenge of automated UI operation, making it easier to avoid being evaluated as a purely business method.
① Draft claims that directly protect product functions. Focus not only on abstract, high-level concepts but also on specific configurations that correspond to your company’s core product functions.
② Articulate the architecture in technical terms. Define the system’s components and processing flow using technical terminology—such as “intermediate representation,” “observation logic,” and “operation commands”—and incorporate them into the claims.
③ For distributed architectures, clearly describe and distinguish “what is done where.” Explicitly define the division of roles among client, server, and agent, and tailor the approach of the claims for each infringing entity.
④ Convert operations into functions to ensure comprehensive coverage. Specify operational terms such as “click,” “type,” and “scroll” in dependent claims and the specification to make design-around difficult.
We’ll assess whether your company’s AI agent is patentable.
Patent attorneys with deep expertise in the IT, software, and AI fields provide comprehensive support—from free assessments of patentability to claim drafting, FTO searches, and filing strategies in Japan, the U.S., and Europe.
Schedule a Free Initial Consultation IT & AI Intellectual Property ServicesQ. What kind of patent is US 12,430,150 B1?
A. It is a U.S. registered patent held by Anthropic that protects a “runtime architecture” enabling an AI agent to automatically operate a multimodal (image + text) user interface.It is a core patent supporting the company’s “Computer Use” (the functionality that allows Claude to operate a computer) and was granted on September 30, 2025. It consists of a total of 20 claims (3 independent and 17 dependent).
Q. How many independent claims does this patent have?
A. There are three. They fall into three categories: Claim 1 (system), Claim 14 (computer-implemented method), and Claim 20 (non-transitory computer-readable storage medium). In line with standard practice for software patents, they are structured to cover different potential infringers.
Q. Why are the “client-side” and “server-side” separated?
A. This architecture is a divided structure in which the client-side handles the construction of agent specifications, the interpretation of intermediate representations, and the translation into execution commands, while the server-side handles the translation of agent specifications into intermediate representations and hosts the agent itself.This configuration, which delegates heavy inference to server-side AI while the client-side handles UI operations, aligns with practical implementation practices; this level of specificity contributes to ensuring patent eligibility.
Q. How can Japanese companies obtain patents for “Computer Use”-type technologies?
A. Rather than focusing on the idea of “AI operating a PC,” it is effective to describe the invention as a specific software architecture, as in this case, including: (1) the division of roles between client and server, (2) an abstraction layer using an intermediate representation, and (3) a multi-stage transformation process involving agent invocation → operation functions → execution commands.
Q. I’m concerned that my work might infringe on this patent. What should I do?
A. This is a registered patent, and the actual scope of rights is determined by the wording of each claim, the doctrine of equivalents, and prior art.Determining whether your company’s agent technology infringes on this patent (FTO = Freedom to Operate analysis) is a specialized task requiring claim interpretation. If you have concerns, please consult a patent attorney well-versed in the IT and software fields to conduct a claim-by-claim comparison.
Important Note Regarding This Article: This article provides a general explanation of technology and legal systems based on published patent applications. While US 12,430,150 B1 is a registered patent, its 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 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 a case-by-case review.The Japanese translation is provided for reference purposes only; the official text is the original English version.