evorix blog

[Patent Attorney's Analysis] Decoding the Claims of Salesforce's Multi-Agent Patent (US2025/0265443) | Practical Examples and Drafting Techniques for AI Agent Patent…

“Coordinating multiple AI agents to solve complex tasks”—in the field of multi-agent AI, which is currently attracting the most attention, a U.S. patent application filed by Salesforce has been published (US2025/0265443 A1, published August 21, 2025, 20 claims in total).In this article, a patent attorney well-versed in AI intellectual property will thoroughly analyze the technical content and patent claim structure, citing the actual claims (scope of protection) directly from the original text.

Furthermore, using this case as a case study, we delve into practical insights on “how to draft claims and specifications for AI agent inventions.” For developers and IP professionals who are thinking, “We want to patent our company’s AI agent, but we don’t know where to start or how to draft it,” this excellent, real-world patent application serves as the best textbook.

💡 Key Point: This article is the “In-Depth Analysis of Individual Patents” installment in our AI Agent Patent Series. Please refer to the “Basics” installment for foundational requirements and the “Case Studies” installment for a comparison of cases from Japan, the U.S., and Europe.

Table of Contents

  1. Key Points of This Patent in 30 Seconds
  2. Basic Patent Information
  3. Technical Explanation | Hierarchical Multi-Agent Architecture
  4. Specific Application Examples (4)
  5. [The Core of This Article] Reading Independent Claim 1 in the Original Text
  6. Analysis of Claim Structure | The “Method/System/Media” Trio
  7. The “Layered Rights” Strategy Seen in Dependent Claims
  8. Reference | How to Draft Claims for AI Agent Inventions
  9. Reference | How to Write a Specification for an AI Agent Invention
  10. How These Will Be Evaluated in Japanese, U.S., and European Examinations
  11. Five Lessons for Your Company’s Applications
  12. Note: This application has been published (under examination)
  13. Frequently Asked Questions (FAQ)

Key Points of This Patent in 30 Seconds

● What is the invention?: A hierarchical multi-agent architecture in which a “manager agent” breaks down complex tasks into smaller tasks, dynamically selects the optimal sub-model, exchanges packages in a format compatible with the target model via an API, and generates the next task based on the output.
● Applicant: Salesforce Inc.(U.S.)
● Scope: 20 claims total (3 independent claims: method, system, and medium)
● Status: U.S. Patent Application Published (A1) = Under examination
● Why it’s important: This application directly seeks to secure rights for the mainstream “manager-worker” pattern of AI agent implementation.

Basic Patent Information

Item Content
Publication Number US 2025/0265443 A1
Application Number 18/738,984
Title of the Invention Systems and Methods for Building Task-Oriented Hierarchical Agent Architectures
Publication Date August 21, 2025
Filing Date June 10, 2024
Priority Date February 19, 2024
Applicant Salesforce Inc.
Inventors Zhiwei Liu and 14 others
Number of Claims 20 (3 independent claims: Claims 1, 11, and 20)
Main Classifications (CPC) G06N3/00, 3/02, 3/04, 3/045 (Neural Networks)
Status Application Published (Under Examination)

Technical Description | Hierarchical Multi-Agent Architecture

According to the abstract, the present invention is a “method for constructing a hierarchical structure of multiple neural network models for executing tasks.” The core processing flow is as follows:

[Receive task instructions] (via data interface) ▼ [First model] — Generate “first subtask” from task instructions ▼ [Select second model] — Dynamically select from multiple models based on the subtask ▼ [Establish API Connection] ── Establish a connection between the first and second models ▼ [Generate Subtask Package] └ Generate in a format compliant with the second model ▼ [Second model executes → Receive output](via API connection) ▼ [First model generates “Second Subtask”] └ Based on the task instruction and the output from the first model

3-tier agent structure

Hierarchy Role (based on specifications)
Manager Agent Communicates with each individual agent and manages task assignments among agents
Sub-Manager Agent An intermediate layer that further breaks down assigned tasks
Individual Agent Execution agents optimized for specific functions

Communication and Optimization Among Agents

Agents exchange “task packages” via an API. Each package contains a prompt tailored to the recipient agent, prompting it to generate an action. Two methods for optimizing each agent are described: (1) fine-tuning the LLM itself, and (2) optimizing the prompts.

Specific Application Examples (4)

Application Examples Description
Online Painter A search agent and a painter agent collaborate to search for visual features online and then generate an image
Interactive Image Understanding Answers human questions based on images through multiple rounds of dialogue
Solving Math Problems A math agent solves equations such as “75 × 34 + 12=” using WolframAlpha integration
Chess Match A chess agent recognizes the board, executes legal moves, and plays against a human

[Key Point of This Article] Read Independent Claim 1 in the Original Text

What is actually protected by a patent is not the abstract or the drawings, but the “claims.” Here, we quote the most important claim in this case—Independent Claim 1 (a method claim)—in its original text.

Claim 1 (Original Text / English)

A method for building a hierarchical structure of a plurality of neural network models for performing a task, the method comprising: receiving, via a data interface, a task instruction; generating, by a first neural network model, a first sub-task from the task instruction; selecting a second neural network model from the plurality of neural network models based on the first sub-task; establishing a first connection, via a first application programming interface (API), between the first neural network model and the second neural network model; generating, by the first neural network model, a first sub-task package in a format compatible with the second neural network model; receiving, via the first connection, a first output from the second neural network model that executes the first sub-task package; generating, by the first neural network model, a second sub-task based on the task instruction and the first output; and causing the task instruction to be jointly performed by one or more selected neural network models from the plurality of neural network models based at least in part on the second sub-task.

Reference Translation by a Patent Attorney (Japanese)

A method for constructing a hierarchical structure of multiple neural network models for executing a task, comprising:
① the step of receiving a task instruction via a data interface;
② the step of generating a first sub-task from said task instruction using a first model;
③ the step of selecting a second model from the plurality of models based on said first sub-task;
④ a step of establishing a first connection between the first model and the second model via a first API;
⑤ a step of generating, using the first model, a first sub-task package in a format compatible with the second model;
(6) a step of receiving a first output from a second model that executes said package via said first connection;
(7) a step of generating a second subtask based on the task instruction and the first output using the first model;
⑧ a step of causing one or more selected models to jointly execute the task instruction based at least in part on the second subtask;

A Clause-by-Clause Analysis of Claim 1

The ingenuity of this claim lies in its breakdown of the abstract idea of “coordinating multiple AIs” into eight specific processing steps. The following limitation is particularly effective from the perspective of patentability (patent eligibility and inventive step):

Limitations Technical Meaning Why It Is Important
④ Establishing an API connection Defines the collaboration between models not as abstract “coordination” but as a concrete communication connection Moving beyond “mere ideas.” Explicitly defining technical implementation
⑤ Generating Packages in Compatible Formats Converting data into a format that the target model can interpret The technical core of agent collaboration. A key point of innovation
③ Dynamic selection based on subtasks Selection based on content rather than a static, fixed configuration The source of intelligence in orchestration
⑦ Generating the next task based on the output Forming a Feedback Loop Autonomy = The Basis of Agent-Like Behavior

💡 Key point: The qualification “a format compliant with the second neural network model” is the technical highlight of this claim.This is not merely a transfer of information, but a technical innovation that enables interoperability between different models, and serves as a strong basis for avoiding rejection on the grounds of an abstract idea (equivalent to U.S. Alice and Japanese Patent Law Article 2, Paragraph 3(k)).

Analysis of Claim Structure | The “Method/System/Media” Trio

In this case, out of a total of 20 claims, the independent claims are organized into three distinct categories. This is a standard practice for software patents.

Claims Category Subject Matter Anticipated Infringers
Claim 1 Method Processing Procedure A person performing the method
Claim 11 System A device equipped with memory and a processor A person who manufactures, sells, or uses the device
Claim 20 Non-transitory machine-readable medium A storage medium on which a program is recorded A person who distributes or provides a program

Claim 11 (System) describes the same processing as Claim 1 in conjunction with the physical components “memory” and “one or more hardware processors.”

Claim 11 (Excerpt) (Original Text / English)

A system ... comprising: a memory that stores the plurality of neural network models and a plurality of processor-executable instructions; a communication interface that receives a task instruction; and one or more hardware processors that read and execute the plurality of processor-executable instructions from the memory to perform operations comprising: ... (The same processing as for Claim 1 follows)

💡 Key Point: Under Japanese examination standards, the question is whether “information processing is specifically implemented using hardware resources.”Claim 11’s explicit mention of “memory” and “hardware processor” is a standard approach to meet this requirement. When filing the same invention in Japan, it is also effective to clearly specify the interaction with hardware resources in system claims.

[Claim Tree (Independent/Dependent Relationships)] ● Claim 1 (Method/Independent) ├─ Claim 2 (Further selection and coordination of the third model) ├─ Claim 3 (Package = Appropriate Prompt) ├─ Claim 4 (Output = Completion Status, etc.) ├─ Claim 5 (Taking Human Instructions into Account) ├─ Claim 6 (Generating a Package in Collaboration with the Fourth Model) │ ├─ Claim 7(Fourth model refines the initial subtask) │ └─ Claim 8 (Tasks are divided and assigned to each model) ├─ Claim 9 (Multi-stage delegation to the fifth and sixth models) └─ Claim 10(Each model is trained independently) ● Claim 11 (System-independent) └─ Claims 12–19 (Dependent claims corresponding to Claims 2–10) ● Claim 20 (Media-independent)

The “Layered Protection” Strategy Seen in Dependent Claims

Even if an independent claim were to be invalidated by prior art, the rights would be maintained as long as the dependent claims survive. Dependent claims function as a “layered defense.” From the dependent claims in this case, we can learn how to skillfully apply limitations.

Claim 3 (Original Text / English)

The method of claim 1, wherein the first sub-task package comprises a first prompt compliant with the second neural network model, instructing the second neural network model to perform the first sub-task.

Claim 3 specifies the contents of the package as acompliant prompt.” By limiting the broader concept of “format” (Claim 1) with the narrower concept of “prompt,” the scope of protection is structured in a tiered manner.

Claim 5 (Original / English)

The method of claim 1, further comprising: receiving ... a human instruction; and generating, by the first neural network model, the first sub-task based on the task instruction and the human instruction.

Claim 5 covers a variation (human-in-the-loop) that incorporates “human instructions” into the process. By covering implementation variations in dependent claims, the patent prevents competitors from circumventing the invention through design changes.

Claim 10 (Original / English)

The method of claim 1, wherein the first neural network and the second neural network are each trained independently.

Claim 10 limits the scope to the point that “each model is trained independently.” It incorporates the actual configuration of combining heterogeneous models into the scope of protection.

💡 Key Points: The standard approach for dependent claims is: (1) stepwise limitation from a higher-level concept to a lower-level concept (Claim 3),② covering implementation variations (human intervention in Claim 5; multi-stage delegation in Claims 6–9), and ③ specifying characteristics of the configuration (independent training in Claim 10).Formulating broad independent claims while using dependent claims to thoroughly cover specific embodiments—this “hierarchy of breadth and narrowness” creates a strong patent portfolio.

Reference | How to Draft Claims for AI Agent Inventions

Based on this case, we present a framework for approaching the drafting of claims for your company’s AI agent inventions. The following is a general reference example for illustrative purposes (these are not the claims from this patent).

📝 Claim Drafting Reference Example (Not the claims of this patent)

[Bad Example | Too Abstract and Prone to Rejection
] “A system that automatically executes a user’s task using multiple AI agents.”

Reason for Rejection: “Executing a task using multiple AI agents” is merely a description of an idea or function, and the technical means are unclear. High risk of rejection under the U.S. Alice standard or Japan’s “invention” requirement.

📝 Claim Drafting Reference Examples (These are not claims of this patent)

[Good Example | Specifies the Processing Flow]
“A receiving unit that receives task instructions; a first model that
decomposes said task instructions into multiple subtasks; a selection unit that selects an execution model from among
multiple execution models based on the content of each subtask;a conversion unit that generates an instruction package converted
into a predetermined format interpretable by the selected execution model; a communication unit
that establishes an API connection with the execution model, transmits the package, and receives execution results; and a control unit that generates the next subtask based on the execution results
; an information processing device comprising the aforementioned units.”

Reason for acceptance: The description is broken down into specific processing steps—receipt → decomposition → selection → conversion → communication → generation of the next subtask—and explicitly states collaboration with hardware and APIs.

Claim Design Checklist: ① Describe the process in terms of specific steps rather than abstract objectives (e.g., “to automate…”); ② Use interfaces between agents (formats, APIs, protocols) as technical hooks; ③ Explicitly state dynamic selection and control logic; /④ Formulate claims in the three categories: methods, systems, and media / ⑤ Cover all implementation variations in dependent claims.

Reference | How to Write a Specification for AI Agent Inventions

Strong claims are only valid when supported by a well-written specification. Below, we outline the particularly important elements of a specification for AI agent inventions, correlating them with the structure of this case.

Element Content to Include Practice in This Case
Technical Problem Clearly define technical challenges, such as the inability of a single agent or LLM to solve complex tasks State explicitly that “a single LLM agent struggles with complex tasks requiring diverse specialized actions”
Solutions Specific architecture and processing flow to resolve the challenges Hierarchical Decomposition + Dynamic Selection + Format Adaptation + API Integration
Implementation Examples (Important) Ensures implementation requirements and support requirements through multiple and diverse examples Four examples: Painter, Image Understanding, Mathematics, and Chess
Definition of Terms Define and specify functional concepts such as “Task Package” and “Adaptive Format” Package = Includes a suitable prompt, etc.
Variations Variations such as human intervention, multi-stage delegation, and independent learning Substantiate variations corresponding to dependent claims in the specification

Pay attention to the requirements for workability and support: AI processing tends to be a “black box.”Abstract statements such as “good results will be obtained by inputting data into the AI” will be rejected in both Japan and the U.S. It is essential to describe in sufficient detail—to the extent that a person skilled in the art can reproduce the process—what input is used, what conditional branching is applied, which model is selected, and into what format the data is converted. The reason this application includes four detailed examples is to meet this requirement.

💡 Key Point: The golden rule of drafting a specification is “broad disclosure to support broad claims.”If you use high-level concepts (such as format or model selection) in the claims, include multiple specific examples (such as prompt formats and selection algorithms) in the specification and incorporate fallback positions for intermediate concepts. This allows you to narrow the claims without adding new subject matter, even if the examiner requests limitations during examination.

How This Is Evaluated in Japanese, U.S., and European Examinations

United States (USPTO) | Alice/Mayo Test

In the U.S., the key question is whether the invention is directed to an abstract idea and, if so, whether it contains an inventive concept. This case is described as a concrete workflow involving technical implementations—such as a set of neural network models, API connections, and a package of compatible formats—and is structured in a way that makes it easy to assert a “technical solution to a technical problem.”This also aligns with the approach outlined in the USPTO’s July 2024 AI Examination Guidelines (Examples 47–49).

Japan (JPO) | Interaction with Hardware Resources

In Japan, the criterion is whether “information processing is specifically implemented using hardware resources.” Given that Claim 11 of this case explicitly mentions “memory” and “hardware processor,” and that specific processing steps—such as data interfaces, API connections, and format conversion—are described, it is considered likely to satisfy patent eligibility as a software-related invention.The key to demonstrating inventive step lies in the technical ingenuity and effects of “dynamic model selection” and “adaptive format conversion.”

Europe (EPO) | Technical Contribution (COMVIK)

In Europe, only features that contribute to the technical nature of the invention are considered in assessing inventive step. In this case, “format conversion enabling interoperability between heterogeneous models” and “coordination via API connections” can easily be positioned as technical solutions to a technical problem (interoperability), making it easier to avoid being classified as a purely business method.

A detailed comparison of examination practices and case examples for AI agent patents in Japan, the U.S., and Europe is provided in “Can AI Agents Be Patented? Patent Cases and Examination Practices in Japan, the U.S., and Europe.”

Five Lessons for Your Own Patent Applications

① Describe the “processing flow” rather than just the “idea.” Break it down into specific steps: reception → decomposition → selection → conversion → connection → output → generation of the next task.

② Use the “interface” between agents as a technical hook. Compatible formats, API connections, and communication protocols are powerful clues that help move beyond abstract concepts.

③ Structure claims into three categories: methods, systems, and media. This captures different infringers and broadens the scope of protection.

④ Use dependent claims to build a multi-layered protection structure. Gradually cover higher-level concepts → lower-level concepts, implementation variations, and configuration characteristics.

⑤ Provide detailed examples of practice. Use diverse application examples to satisfy the requirements for practicability and support, thereby underpinning broad claims.

Note: This application is published (under examination)

The scope of protection is not yet determined: This application (US2025/0265443 A1) is a published U.S. patent application and is not a registered patent. The final scope of protection will be determined following future examination (including responses to office actions and amendments).This article is a general explanation of the technology and system based on published application documents and does not guarantee any specific scope of protection or validity. When making business decisions (such as infringement analysis or FTO assessments), third parties must always refer to the latest status information, the official USPTO transcript, and individual expert reviews.

Regarding the source of the claim citations: The original text of the claims in this article is based on published application data available on FreePatentsOnline. For legally significant purposes (such as invalidity, infringement, FTO analysis, or filing an application), please verify the exact wording in the official PDF from the USPTO Patent Public Search.The Japanese translation is a reference translation provided by a patent attorney for the purpose of understanding; the official text is the original English version.

We’ll assess whether your company’s AI agent technology is patentable.

Patent attorneys with expertise in the IT and software fields provide comprehensive support—from technical consultations and free assessments of patentability to claim drafting and filing strategies in Japan, the U.S., and Europe.

Schedule a Free Initial Consultation IT & AI Intellectual Property Services

Frequently Asked Questions (FAQ)

Q. What kind of patent is US2025/0265443 A1?

A. It is a U.S. patent application by Salesforce regarding a “Task-Oriented Hierarchical Agent Architecture” (published August 21, 2025; 20 claims in total).It covers foundational technology for multi-agent AI, in which a higher-level “manager agent” breaks down complex tasks into subtasks, dynamically selects the optimal lower-level model for each subtask, and passes a package in a format compatible with the target model via API to enable collaborative processing.

Q. Is this a registered patent?

A. No. The “A1” at the end indicates a published U.S. patent application, and as of 2026, it is pending examination. The final scope of the patent rights will be determined during the upcoming examination.

Q. How many independent claims does this patent have?

A. There are three. They fall into three categories: Claim 1 (method), Claim 11 (system), and Claim 20 (non-transitory machine-readable medium), all of which share the same core limitations.The remaining claims 2–10 and 12–19 are dependent claims.

Q. Can a multi-agent mechanism be patented?

A. The idea of “using multiple AIs” is abstract in itself and difficult to patent; however, if it is described as a specific information processing flow—such as in this case, “breaking down tasks, generating packages in a suitable format, establishing API connections to receive output, and generating the next task”—the likelihood of obtaining patent protection increases in both Japan and the United States.

Q. Why are there three types of claims—method, system, and medium?

A. To broaden the scope of protection.Method claims protect processing procedures, system claims protect devices (such as memory and processors), and media claims protect storage media on which programs are recorded. This allows us to cover different types of infringers—such as “those who implement the invention,” “those who manufacture or sell the device,” and “those who distribute the program”—all based on the same invention.

Q. How can Japanese companies obtain patents for multi-agent technology?

A. It is effective to focus the claims on technically definable elements, such as communication protocols between agents, dynamic selection logic, and format conversion of task packages. The key is to describe these as system processing flows, rather than limiting the claims to mere business ideas or creative use of prompts.

Q. What should be particularly noted in the specification for an AI agent invention?

A. There are three key points: (1) clearly link the technical problem (e.g., a single agent cannot solve complex tasks) to the solution; (2) describe multiple and diverse examples to satisfy the requirements for practicability and support; and (3) translate functional concepts—such as “compatible formats” and “API connections”—into concrete processing steps.

Q. Can I still obtain a patent for technology that was disclosed prior to filing the application?

A. As a general rule, once technology has been disclosed in academic papers, open-source software (OSS), press releases, or similar venues, it loses novelty and cannot be patented. While Japan has an exception to the loss of novelty under certain conditions (Article 30), the requirements vary by country and this exception is not universal; therefore, the golden rule is to complete the patent application before any public disclosure.

Recommended Reading (AI Agent Patent Series):Can
AI Agent Technology Be Patented? Examination Standards and Key Points for Obtaining a Patent (Basics)Can
AI Agents Be Patented? Patent Cases and Examination Practices in Japan, the U.S., and Europe (Case Studies)

Source