This application introduces technologies such as the proprietary “Adept Workflow Language” developed by Adept, a three-tier prompt system consisting of clicks, steps, and tasks, and DOM abstraction.Furthermore, since this is a “published application (A1)” rather than a granted patent, we’ll also cover essential practical topics for practitioners, such as “how to read published applications” and “what deleted claims reveal.”
💡 Key Point: This article is the final installment of the AI Agent Patent Series: Anthropic Edition. Please read it in conjunction with the Strategic Analysis, the US 12,430,150 analysis, and the US 12,566,913 analysis.
Table of Contents
| Item | Content |
|---|---|
| Publication Number | US 2025/0299023 A1 |
| Title of the Invention | Systems and Methods for Configuring Artificial Intelligence Agents to Automate Multimodal Interface Workflows |
| Publication Date | September 25, 2025 |
| Filing Date | October 8, 2024 |
| Priority Date | March 20, 2024 |
| Current Applicant | Anthropic PBC (transferred from Adept AI Labs Inc. on April 17, 2025) |
| Status | Published Application (Under Examination) |
| Related | One of Anthropic’s “Computer Use” patent families |
Before reading this article, make sure you understand the “kind code” at the end of a patent number. This is a fundamental concept in patent practice and will also be helpful when you come across patents in the news.
| Symbols | Meaning | Status of the Rights |
|---|---|---|
| A1 | Published application | Under examination; status undetermined. Rights have not yet been granted |
| B1 | Registered Patent (including those that have not undergone publication, etc.) | Rights have been granted |
| B2 | Registered patent (registered after publication of the application) | Patent Granted |
What to Keep in Mind When Reading “A1”: In the case of US 2025/0299023, “A1” refers to the application publication, and it has not yet been granted as a patent. The claims may be amended during the examination process, and the final scope of protection is subject to change.This is fundamentally different from the US 12,430,150 B1 and US 12,566,913 B2 discussed in the previous two articles, which were already registered (B1/B2).
💡 Key Point: A1 (published application) is a “preview of what may become a right in the future,” while B1/B2 (registered patents) are “established rights.” Please keep in mind that even within the same technology family, the meaning and interpretation differ completely depending on the status.
Anthropic’s “Computer Use” is protected in a layered manner across multiple patents covering different layers. Let’s organize the relationships among these three patents, including this one.
| Layer | Patent/Application | What Is Protected | Example |
|---|---|---|---|
| ① Architecture | US 2025/0299023 A1 (this article) | Mechanism for Creating Agents / Dedicated Language | “Writing a Script” |
| ② Execution Platform | US 12,430,150 B1 | Runtime (Client/Server Division) | “Stage” |
| ③ Agent Body | US 12,566,913 B2 | Agent Intelligence and Behavior (DSL, etc.) | "Actors" |
💡 Key Point: Dividing the product into three layers—“Script (Construction) → Stage (Execution Platform) → Actor (Agent)”—and filing the patent application accordingly is a standard strategy for building a strong patent portfolio. This structure ensures that even if a competitor manages to circumvent one layer, they may still infringe on another.
The specification of this patent application describes “Adept Workflow Language (AWL)” as “an expressive custom language that allows users to easily describe powerful multimodal web operations (on Adept’s model).”
In other words, it is a dedicated programming language for humans to “author” the behavior of AI agents. Whereas the DSL described in US 12,566,913—which we covered previously—was “the language used by agents at runtime,” this language is positioned as “the language for humans to design workflows,” thereby supporting the upstream “construction” phase.
Claim 1 stipulates that the agent specification function constructs “multi-tiered prompts at the click level, step level, and task level.”
By organizing prompts into three levels of granularity—from abstract tasks to specific actions—complex workflows can be handled in a structured manner. This hierarchical structure itself functions as a claim limitation, thereby defining the scope of the rights.
Another technical highlight is the limitation that the agent specification function “conceals the complexities of a webpage’s DOM (Document Object Model) by abstracting interface elements into high-level references.”
Web pages have a vast and complex DOM structure internally. Rather than having the AI handle this structure directly, abstracting it into high-level references such as “this button” or “this input field” makes the creation and execution of agents easier and more robust.
💡 Key Point: “Concealing DOM complexities” is precisely a technical solution to a technical challenge (complex web structures). Such descriptions serve as strong evidence that this is “not an abstract idea” in the U.S. Alice case, Japanese software-related inventions, and the European COMVIK case.
The specification further mentions a mechanism that “intercepts user operations performed on the interface, translates them into activation commands, and generates a training dataset.”
Recording how a human actually operates the software and converting that into training data for an AI agent—this is, in essence, an “show-and-teach” approach. The fact that the scope of this application extends to include the method of creating training data when “building” the agent demonstrates the breadth of its scope.
US 2025/0299023 A1 | Claim 1 (Original Text / English)
A system for configuring an agent to automate multimodal interface workflows, comprising: a memory; one or more processors coupled to the memory and configured to: construct agent specification functions using prompts and agent functions, wherein the agent specifications are configured to automate a multimodal interface workflow, and wherein the agent specification functions are configured to construct multiple levels of prompts, including click-level prompts, step-level prompts, and task-level prompts, and wherein the agent specification functions conceal the complexities of a web page document object model (DOM) by abstracting interface elements into higher-level references; and generate multimodal agent calls—by translating the agent specification functions—that cause an agent to implement the agent functions to produce outputs that perform tasks corresponding to the prompts, wherein the agent calls trigger one or more machine-actuated actions based on the outputs as synthetic actions that automate the multimodal interface workflow, wherein automating the multimodal interface workflow includes modifying elements of the interface, and wherein the interface is a user interface.
Reference Translation by a Patent Attorney (Japanese)
| Limitations | Technical Meaning | Reason for Inclusion |
|---|---|---|
| Three Levels of the Prompt | Granularity Management: Task → Step → Click | Concretization of Control Structures |
| Hiding (Abstracting) DOM Complexity | Mapping Complex Web Structures to High-Level References | Technical Solutions to Technical Challenges |
| Translation of Specification Functions to Agent Calls | Converting Design into Executable Calls | Concrete Processing Bridging Construction and Execution |
A review of the published application documents reveals that claims 6, 11, and 13 are marked as “canceled.” This is evidence that the applicant made amendments after filing the application.
Claim cancellations are common during the process of responding to examination or refining the scope of protection. When reading a published application, it is important to distinguish between these “active claims” and “canceled claims” to accurately understand what is currently being claimed.
💡 Key Point: While registered patents (B1/B2) are snapshots of “finalized rights,” published applications (A1) represent the “interim status of the examination process.”Claim deletions and amendments occur routinely, and the scope of protection evolves. This is precisely why it is essential to continuously monitor the latest case history (file wrapper) during competitive analysis and FTO evaluations.
Another key point to note in this application is the record of the assignee’s transfer. This application was originally filed by Adept AI Labs Inc., but was reassigned to Anthropic PBC on April 17, 2025.
① “The creation process” itself can be patented. Not only the technology that executes a product, but also the mechanisms used to build and author it (such as specialized languages and design tools) are eligible for patent protection.
② Use abstraction and hierarchization as technical hooks. Structural innovations that manage complexity—such as “concealing the complexity of the DOM” or “three-tier prompts”—can serve as strong limitations in patent claims.
③ Include the methods for creating training data within the scope of protection. Consider protecting data pipelines as well, such as the process of converting user operation logs into training data.
④ Utilize patent assignment records. Trends among competitors and potential acquisition targets can be gleaned from the applicant’s assignment history.
⑤ Monitor “progress information” for published applications. Since the scope of rights for applications at the A1 stage is subject to change, continuous monitoring is essential if you have related business operations.
Why not secure intellectual property rights for your company’s AI agents and the underlying development technologies?
Patent attorneys with expertise in the IT, software, and AI fields provide comprehensive support—from free feasibility assessments and claim drafting to monitoring competing patents, 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 2025/0299023 A1?
A. This is a published U.S. patent application by Anthropic that covers a system and method for “configuring” an AI agent that automates a multimodal interface.Its key features include constructing “agent specifications” based on prompts and agent functions, assembling prompts across three hierarchical levels (clicks, steps, and tasks), and abstracting the complexity of a web page’s DOM. Published on September 25, 2025; currently under examination.
Q. What is the difference between “A1,” “B1,” and “B2”?
A. The symbol at the end of a U.S. patent number indicates its status. “A1” denotes a published application—it is still under examination and the rights have not yet been established. “B1” and “B2” denote granted patents—the rights have already been established.Since this A1 application is currently under examination, the final scope of rights may change as a result of future amendments and examination.
Q. What is Adept Workflow Language (AWL)?
A. It is an “expressive custom language” described in the specification that allows users to easily describe multimodal web operations. This technology was acquired by Anthropic from Adept and serves as a dedicated language for humans to “author” workflows for AI agents.
Q. Why does the name “Adept” appear in an Anthropic patent?
A. This patent application was originally filed by Adept AI Labs, but the applicant was reassigned to Anthropic PBC on April 17, 2025.The fact that Anthropic acquired Adept’s team and technology through an “acqui-hire” in 2024 is corroborated by the record of the assignee transfer.
Q. What should competitors be aware of at the published application stage?
A. Published applications may be registered in the future, and if registered, they may give rise to claims for certain compensation retroactive to the publication date (in the case of Japan).Even for U.S. applications, the scope of protection may change due to amendments, so if you are engaged in related business activities, it is important to monitor updates and, if necessary, consult with experts from a Freedom to Operate (FTO) perspective.
Important Note Regarding This Article: This article provides a general explanation of the technology and system based on published patent application bulletins. US 2025/0299023 A1 is a published application currently under examination and has not yet been granted. Since the claims may be amended, the final scope of protection has not yet been determined.The cited claims, abstract, and description are based on published application data (Google Patents, FreePatentsOnline, etc.); however, for legally significant purposes (FTO, infringement analysis, filing applications, etc.), please be sure to verify the USPTO official text and the latest status information, and consult with an expert for a case-by-case review.The Japanese translation is provided for reference purposes only; the official text is the original English version.