Understanding How to Patent Agentic AI Systems

Introduction

Artificial intelligence is evolving beyond simple pattern recognition and content generation into autonomous decision-making. Agentic AI systems act independently—or with limited human supervision—to achieve specific objectives. These systems employ various machine learning techniques, including reinforcement learning, natural language processing, and knowledge representation, to operate effectively in dynamic environments.

Agentic AI is fundamentally different from Generative AI (GenAI) systems like ChatGPT or Grok. While GenAI specializes in creating content based on human prompts, Agentic AI systems execute sequences of actions autonomously to accomplish predefined goals. Unlike GenAI, which requires human input at each step, Agentic AI can act continuously, adjusting its behavior based on environmental feedback.

How Agentic AI Systems Work

Agentic AI systems function by executing actions or sequences of actions that align with an objective function—a set of rules that define success for the system. These systems process data inputs, assess their current state, and take actions to maximize or minimize an outcome based on this objective function.

At any given moment, an Agentic AI system relies on a percept, which represents its current understanding of the environment and its goal. The agent then selects and executes an action based on its percept. Crucially, the percept can change dynamically as the agent interacts with its surroundings, allowing the system to adapt and improve its decision-making over time.

The objective function can be explicitly defined—for example, in a robotic system designed to maximize energy efficiency—or it can be learned over time through reinforcement learning. Some objective functions may take the form of:

• Reward functions, where the AI agent seeks to maximize positive outcomes.
• Cost functions, where the AI minimizes penalties or undesirable results.

By continuously adjusting actions based on percepts and the objective function, Agentic AI systems can perform complex, autonomous decision-making across diverse applications.

Drafting Patent Claims for Agentic AI Systems

As Agentic AI technology advances, intellectual property protection becomes essential. Patents can cover different aspects of these systems, including their structure, training processes, and application in specific domains. Below are two notable examples of patent claims related to Agentic AI:

Building an Agentic AI System (Example: US 12,111,859)

This patent describes an agentic system that uses an orchestrator to supervise multiple subordinate agents. 

The orchestrator uses multimodal models (models that can interact with different types of data, such as text, pictures, and video) to “process or deconstruct” a prompt into a series of instructions for the agents. In turn, each agent uses one or more machine learning models to process inputs derived from the deconstructed prompt. 

The orchestrator requests information from the agents, which provide the orchestrator with data responsive to the request. The orchestrator analyzes the data to generate a response to the prompt. If the response isn’t sufficient, the orchestrator requests the agents for additional data. The orchestrator can provide additional instructions, including follow-up questions, to the agents to aid them in providing additional information to the orchestrator. The agents can use machine learning models of their own to generate the information consumed by the orchestrator. This process continues until the orchestrator can generate a sufficient response. 

The dependent claims discuss the types of data processed (time series data, structured data, and unstructured data), operations performed (calculation, translation, formatting, and visualization), and training (the agents are trained on different domain-specific machine-learning models). 

Key Takeaways for Patent Drafting:

• Clearly describe the system structure. Is there specialized hardware or software? How do system components interact with one another? 
  • “wherein at least one agent employs a type system to unify incompatible data from disparate data sources”
• Define the roles and interactions between the different agents. What is the objective of the system and how do the agents work together to achieve it? Is there a hierarchy? What types of functions do agents perform? How many functions can each agent perform?
  • “each agent employs one or more machine-learning models to process disparate inputs or different portions of an input associated with the prompt”
  • “the orchestrator generates intermediate instructions associated with the additional retrieval requests to the plurality of agents”
  • “outputting… a validated response of the one or more responses to the input that satisfies context validation criteria and a portion of data retrieved by the one or more agents related to the input”
  • “the context validation criteria includes a threshold for identifying source material from an enterprise data system that corroborate the response”
  • “managing the plurality of agents comprises iterative processing or multiple instructions from the orchestrator”
• Explain how data flows within the system. Are there intermediate processing steps? What types of data can be input to the system? What kind of data is output from the system?
  • “the orchestrator employs one or more multimodal models to process or deconstruct a prompt into a series of instructions for different agents”
  • “the orchestrator provides additional retrieval requests to the one or more agents to retrieve additional data to satisfy a context validation criteria associated with the input”
  •  “outputting the portion of data… includes a source citation for the at least a portion of the validated response”
  • “retrieving the data from multiple data domains includes time series data, structured data, and unstructured data”

Training an Agentic AI System (Example: US 11,625,314)

This patent focuses on training an agentic system to mimic human users in interacting with a software application. An AI agent learns how to use a first version of the software by using machine learning techniques to process historical session logs generated by users of the software product. The agent then logs into the software and performs a first task on a first version of the software product, as it has been trained to do. The agent then logs into a second version of the software with different features and either perform the first task or a second task on this version of the software. A report is generated, highlighting the performance of the AI agent in using the updated software. 

Dependent claims discuss the types of machine learning techniques used (pattern recognition or classification models), how the AI agent can use a regression model to identify why error messages can occur, how the AI agent can learn how to respond to the software as a person of a particular demographic would, and how, based on the report, the software product is modified to improve usability. 

Key Takeaways for Patent Drafting:

• Define the training process in detail, including process steps and their sequence. How many training stages are there? Is pre-processing performed? Example 39 of the 2019 Patent Eligibility Guidance may be a useful guideline for drafting subject-matter eligible claims.
  • (1) Instructing… the AI agent to apply machine learning techniques to a dataset to learn to use a first version of a software product…; (2) providing to the AI agent… login credentials that grant the AI agent access to the software product; (3) instructing… the AI agent to use the login credentials to log into and use the first version of the software product to perform a first task; (4) determining, by the computing device, that the AI agent successfully performed the first task using the first version of the software product; (5) instructing… the AI agent to use the login credentials to log into and use a second version of the software product to perform at least one of the first task or a second task, wherein the second version of the software product includes one or more new or modified features… (6) generating… a report based on the AI agent logging into and using the second version of the software product to perform the first task or the second task
• Specify the input data and expected output of the trained model. What types of data (e.g., images, video, or text) are processed and/or output by the system? Are there intermediate data products that are used in training?
  • “the dataset is based on data in historical session logs that include records of multiple users logging into and using the first version of the software product”
  • “generating… a report based on the AI agent logging into and using the second version of the software product to perform the first task or the second task”
  • “the dataset is generated using only records, in the historical session logs, of users with a shared demographic characteristic such that the AI agent learns to use the software product as a user with the demographic characteristic”
  • “generating the report comprises comparing, by the computing device, the historical session logs with an AI session log of the AI agent logging into and using the second version of the software product to perform the first task or the second task.”
• Describe the actions the agent is trained to perform. Detail any instructions provided to the agent during training. Does the agent interact with a user interface? Does the agent access the Internet? Can the agent perform any debugging? How does the agent act in case of an error or unexpected behavior? Does the agent utilize a generative AI system? 
  • “instructing, by the computing device, the AI agent to apply machine learning techniques to a dataset to learn to use a first version of a software product”
  • “instructing, by the computing device, the AI agent to use the login credentials to log into and use the first version of the software product to perform a first task”
  • “instructing, by the computing device, the AI agent to use the login credentials to log into and use a second version of the software product to perform at least one of the first task or a second task”
  • “the AI agent applies a regression model to the dataset to identify causal factors for one or more error messages that could be received while using the software product”
  • “the AI agent applies one or both of a pattern recognition model and a classification model to the dataset to recognize normal patterns of user behavior and identify unusual or outlier behaviors while using the software product”
  • “the AI agent applies a decisioning model to the dataset to identify actions suited to achieving particular tasks based on available options while using the software product”
• Explain how the system monitors and evaluates the agent’s learning. Does the agent generate media (e.g., text files or images) of actions it is taking? 
  • “determining, by the computing device, that the AI agent successfully performed the first task using the first version of the software product”

Technological Use Cases of Agentic AI

Agentic AI systems have broad applications across industries. Some notable use cases include:

• Autonomous Control (Vehicles & Robotics): AI-powered systems manage self-driving cars, industrial robots, and unmanned aerial vehicles, making real-time decisions for navigation and safety.
• Customer Support: AI agents provide automated customer service, handling inquiries, troubleshooting issues, and escalating complex cases to human representatives.
• Healthcare: AI-driven systems assist in diagnosis, treatment recommendations, and patient monitoring, improving clinical outcomes.
• Workflow Management: AI agents optimize business processes, coordinating tasks, resources, and schedules to enhance efficiency.
• Financial Risk Management & Investment Decision-Making: AI-powered trading systems analyze market data and execute trades based on predefined investment strategies.
• Content Creation: While GenAI creates content, Agentic AI can autonomously plan, schedule, and distribute it, adapting based on user engagement metrics.

Conclusion

Agentic AI represents a paradigm shift in artificial intelligence, moving beyond content generation to autonomous decision-making and action execution. These systems leverage advanced machine learning techniques to adapt to changing environments and achieve goals with minimal human oversight. As the technology advances, patent protection will play a crucial role in defining and securing innovations in Agentic AI, ensuring companies can leverage these systems across a wide range of industries.

Subscribe To Viewpoints