In this series so far, we have discussed the construction (US 2025/0299023), the execution infrastructure (US 12,430,150), and the agent itself (US 12,566,913).This patent protects the fourth layer—the “data flow” from training to inference (the MLOps infrastructure). A patent attorney specializing in AI intellectual property will explain it while citing the actual claims.
💡 Key Point: This article is the fourth in a series of individual analyses of Anthropic’s patents. Reading it alongside the Strategic Analysis edition will provide a comprehensive, three-dimensional understanding of Anthropic’s patent network surrounding “Computer Use.”
Table of Contents
| Item | Content |
|---|---|
| Publication Number | US 2025/0299074 A1 |
| Title of the Invention | Data Flow Logic for Providing Artificial Intelligence Agents that Automate Multimodal Software Usage |
| Publication Date | September 25, 2025 |
| Filing Date | October 8, 2024 |
| Priority Date | March 20, 2024 |
| Applicant | Anthropic PBC (amended on April 17, 2025) |
| Number of Claims | 20 |
| Status | Published Application (Under Examination) |
Anthropic is providing multi-layered protection for a single product, “Computer Use,” across four distinct layers. Let’s take a look at the big picture, including this application.
| Layer | Patents/Applications | What Is Protected | Example |
|---|---|---|---|
| ① Construction | US 2025/0299023 A1 | Mechanism for Creating Agents · Dedicated Language | Script |
| ② Execution Platform | US 12,430,150 B1 | Runtime (Client/Server) | Stage |
| ③ Agent Body | US 12,566,913 B2 | Agent Intelligence and Behavior | Actor |
| ④ Data Flow | US 2025/0299074 A1 (this article) | Learning-to-Inference Pipeline and Training Data | Blood Flow and Nutrition |
💡 Key Point: In addition to the “visible elements” of “script, stage, and actors,” this case covers “blood flow”—that is, data and the pipeline. Since an AI’s competitiveness is determined by its data infrastructure, securing rights to this fourth layer is a crucial move that adds depth to the patent portfolio.
The core of Claim 1 lies in the clear separation of these two types of servers.
In machine learning systems, it is common practice to separate the “environment for training models” from the “environment for actually serving users (inference/production).” This case claims the very basic structure of MLOps (Machine Learning Operations) itself within the context of AI agents.
“Data flow logic” connects these two server clusters and manages the flow of data. According to Claim 1, this logic performs the following four functions:
| Stage | Data Flow Logic Operations |
|---|---|
| During Training | Provides agents and training datasets to the training server for training |
| Deployment | Configure the production server with the trained agent |
| Inference | Provide the client’s prompt to the production server and have it translated into an agent call |
| Response | Make the generated output available to the client |
💡 Key Point: By establishing an overarching component called “Data Flow Logic,” the entire lifecycle—from training to deployment to inference to return—is treated as a single system. This is a system-patent-like configuration that focuses on the “flow” rather than individual components.
The most strategically important aspect of this case is the content of the training data specified in Claim 1.The claim specifies that the training dataset includes “images of predefined agentic trajectories for multimodal interface automation tasks,” and that “each trajectory is accompanied by detailed multimodal annotations describing the agent’s actions and decisions.”
In other words, this is high-quality training data that records and annotates the sequence of actions taken by an AI agent—specifically, “what it saw on the screen, how it reasoned, and how it interacted with it.” Such data cannot be created overnight.Proprietary training data serves as a “moat” that competitors cannot easily replicate, and Anthropic’s strategy is evident in the fact that it has incorporated this into the elements of its claims.
US 2025/0299074 A1 | Claim 1 (Original Text / English)
A system for providing artificial intelligence agents that automate software usage, comprising: a plurality of training servers configured to train agents during training; a plurality of production servers configured to execute the trained agents during inference; a plurality of training datasets, wherein the plurality of datasets includes images of defined agentic trajectories for multimodal interface automation task workflows, each trajectory accompanied by detailed multimodal annotations describing actions and agentic decisions; and a data flow logic configured to: during training, provide the agents and the plurality of training datasets to the training servers to cause the training servers to train the agents on the plurality of training datasets and thereby produce the trained agents; configure the production servers with the trained agents for use during inference; during inference, provide multimodal interface automation prompts issued by clients to the production servers to cause the production servers to translate the prompts into multimodal interface automation agent calls, wherein each multimodal interface automation agent call specifies interface actions to cause the trained agents to generate outputs that automate multimodal interface workflows in response to the multimodal prompts; and make the outputs available to the clients.
Reference Translation by a Patent Attorney (Japanese)
| Limitations | Technical Meaning | Reason for Effectiveness |
|---|---|---|
| Separation of Training Server and Production Server | MLOps Architecture for Training ↔ Inference | Specific Hardware Configuration of the System |
| Annotated Agent Trajectories (Training Data) | High-Quality Data Recording Actions and Decisions | Data Mine: Securing Rights to the Source of Differentiation |
| Data Flow and Logic | Oversight of the Entire Lifecycle | The Core of System Patents: Capturing the “Flow” |
Claim 1 of this case claims not a specific small component, but rather the “entire system,” which includes the training server, production server, training data, data flow, and logic. This has strategic significance.
① It is difficult to circumvent. Businesses that provide AI agents on a large scale inevitably maintain some form of “training environment” and “production environment” and transfer data between them. By covering the entire system, the scope of protection is more likely to encompass variations even if some components are altered.
② Incorporating intangible assets—such as data—into the scope of protection. By making “annotated trajectories” a claim element, the patent brings data assets—which are difficult to imitate—within the scope of legal protection.
Note: Broad claims carry corresponding risks: Claims that broadly cover the entire system also carry a correspondingly higher risk of being invalidated by prior art (particularly since the general MLOps architecture is a field with a great deal of prior art). We must closely monitor future developments to see to what extent this application will be granted during examination (this case is currently pending).Striking a balance between broad and narrow claims is standard practice.
Because this application includes a specific system configuration—comprising a training server and a production server—along with the technical element of annotated data, it is well-positioned to avoid being classified as an “abstract idea” under the Alice/Mayo tests. However, since there are many prior art examples of MLOps architectures, inventive step (novelty and non-obviousness) could become a point of contention.
Since information processing—involving server clusters, data flows, and logic—is concretely implemented using hardware resources, this configuration is likely to satisfy patent eligibility as a software-related invention. The key to demonstrating inventive step lies in the data design of “annotated agent trajectories” and the technical ingenuity involved in managing data flows.
While it is relatively easy to argue for a contribution to the technical system of the learning-to-inference pipeline, it is effective to explicitly state technical effects (such as efficiency and reliability) to avoid the invention being evaluated as a purely “operational arrangement.”
① Protect not only the product but also the “infrastructure.” Not only the technology being executed but also the learning-to-inference pipeline (MLOps infrastructure) is eligible for patent protection.
② Incorporate proprietary training data into your claims. Methods and structures for creating data that is difficult to imitate—such as “annotated trajectories”—serve as a powerful “moat.” Consider incorporating them into claim elements.
③ File applications covering both the entire system and its components. Balance the risk of invalidation with the difficulty of circumvention by combining broad system claims with narrow, robust component claims.
④ File applications by dividing the product into multiple layers. Secure intellectual property rights for each layer—such as the architecture, execution platform, main body, and data infrastructure—to build a robust patent portfolio.
Why not protect not only your company’s AI and products, but also your “data and pipelines”?
Patent attorneys with deep expertise in the IT, software, and AI fields provide comprehensive support—from free feasibility assessments to claim drafting (including data and MLOps), monitoring of competing patents, 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 2025/0299074 A1?
A. It is a U.S. patent application publication by Anthropic that covers “data flow logic” for providing AI agents that automate software usage.A key feature is the separation of the training server (which trains the agent) and the production server (which performs inference using the trained agent), with the system managing the data flow between them (supply of training data, deployment of trained models, processing of client prompts, and return of output).Published on September 25, 2025; currently under examination; 20 claims in total.
Q. What is “Data Flow Logic”?
A. It is a mechanism that manages how data flows within a system. In this application, it handles a series of “data traffic control” tasks: (1) supplying training data to the server during training, (2) deploying trained agents to the production server, (3) routing client prompts to the production server during inference, and (4) returning outputs to the client.
Q. How is this patent different from the previous three?
A. Anthropic’s “Computer Use” is protected across four layers: ① Architecture (US 2025/0299023), ② Execution Platform (US 12,430,150),③ the agent itself (US 12,566,913), and ④ this patent covers “data flow = the pipeline from training to inference (MLOps infrastructure).” It’s a structure that covers not only the product’s “skeleton” but also its “blood flow.”
Q. What are “annotated agent trajectories”?
A. This is the core of the training data described in Claim 1. It consists of “images of agentic trajectories” from multimodal interface automation tasks, annotated with “detailed descriptions of the agent’s actions and decisions.”This high-quality training data records “how the AI agent thought and acted,” serving as a source of competitive advantage (a “data moat”).
Q. Are there any points that could serve as a reference for our company’s AI development?
A. The key point is that not only the “product itself” but also the “data pipeline from training to inference” and “methods for creating proprietary training data” may be eligible for patent protection. In particular, mechanisms for generating proprietary training data (such as annotated trajectories) serve as a “moat” that competitors cannot easily replicate, making this an area highly worthy of consideration for patent protection.
Important Note Regarding This Article: This article provides a general explanation of technology and the patent system based on published patent application bulletins. US 2025/0299074 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 specification are based on published application data (Google Patents, FreePatentsOnline, etc.); however, for legally significant purposes (FTO, infringement analysis, filing, etc.), please be sure to verify the USPTO official text and the latest prosecution history, 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.