Methods and systems for extending the capabilities of an ai agent
Patent Information
- Application Number
- PCT/EP2025/060479
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2026-10-22
Smart Images
Figure EP2025060479_22102026_PF_FP_ABST
Abstract
Description
[0001] Methods and Systems for Extending the Capabilities of an Al Agent
[0002] TECHNICAL FIELD
[0003] This disclosure relates to methods and systems for extending the capabilities of an Al agent.
[0004] BACKGROUND
[0005] There is a fundamental trade-off between the Al model’s complexity and its universality across different tasks and domains -the more an Al model is versatile, the more complex it tends to be. These challenges can prevent successful deployment of Al agents in general and in a mobile network, for example, specifically.
[0006] A natural approach to circumvent the above challenges is to rely on multiple small Al agents specialized in pre-defined tasks and narrow domains, instead of a one big Al agent capable of generalizing across different tasks and domains. Each such Al agent would be small enough to run efficiently using little resources, and at the same time be capable of providing accurate solutions to end users, provided that the users’ queries match the area of the agent’s specialization.
[0007] However, model specialization also has limitations. One is that specialized agents perform poorly at the edges of their specialization domains, and they often fail to generalize. It could be significant problem when the agent’s environment evolves or when the user’s queries change their characteristics, for example due to seasonal traffic or weather changes in a large city. To tackle this issue and to provide seamless coverage of intelligent services, the next-generation mobile network must incorporate mechanisms for automatic update of these specialized Al agents.
[0008] Fig. 1 presents a setting in which an end user sends a query to a specialized agent (left side). If the query fits well to the agent’s specialization, the agent solves the query and responds with a solution. However, if the query does not fit well to the agent’s current knowledge or if the agent is unsure about the accuracy of the solution, then the agent should evolve to become capable of reliably solving the received query and also similar new queries in the future (right side). Thus, the technical problem addressed by this invention is how to extend the capabilities of an agent in order to solve new types of queries posed by end users.
[0009] There exist several methods for extending capabilities of Al agents, listed below:
[0010] a) Model re-training using new training data
[0011] This method involves re-training or fine-tuning the agent using new training data. The new training data should cover new scenarios compared with an initial training dataset, so that at the end of re-training the agent is capable of solving new queries. This solution has several drawbacks, including a need to collect or generate new training data, perform costly and timeconsuming training, and performance validation. Furthermore, there exists a possibility that the re-trained agent loses some of its previous capabilities while learning the new ones. Finally, this solution is not adapted to situations in which the environment evolves constantly, which would require constant re-training.
[0012] b) Augmenting the model with pre-trained sub-modules
[0013] This method involves incorporating or appending new ML-submodules trained to perform new specialized tasks, e.g. detecting objects in snowy conditions, to the core ML model of the Al agent. This method enables augmenting capabilities of Al agents without the need of re-training using new training data. However, it is difficult to implement in practice due to compatibility issues between the core model and sub-modules. Furthermore, in the context of constantly evolving environment, the augmented model would become ever bigger and resource-consuming up to the point if impracticality.c) In-context learning using human-crafted prompts
[0014] This method takes advantage from the observation that language models that drive the modem Al agents can produce new or better outputs when being prompted. A prompt is usually a textual message appended before the input query, and which guides the inferred model to produce a desired output. Currently, such prompts are usually hand-crafted by humans by trials-and-errors, which is a very slow and inefficient process.
[0015] d) Automated prompt generation
[0016] In the view of costly human-crafted prompting, there exist increasing efforts to automate prompt generation for Al agents using another foundation model fine-tuned to produce prompts. The first results show that automatic prompt generation is a very promising direction, and allows Al agents to solve new tasks without re-training or augmenting their core ML models. However, the existing solutions focus on automatic prompt generation for a specific query, not a class of similar queries. It means that this method in its current form does not allow to practically extend the capabilities of an Al agent, because the generated prompt cannot be generalized or re-used to future queries. Furthermore, most of the solutions enable automatic prompt generation for one particular Al agent. In order to become practical, the foundation model producing prompts should be capable of producing prompts for a wide range of Al agents specialized over different domains, and having own requirements, knowledge, and characteristics.
[0017] This solution overcomes all above limitations, in particular those mentioned in point d). Furthermore, it takes into account unique characteristics and constraints of mobile networks.
[0018] SUMMARY
[0019] The present disclosure attempts to solve these problems.
[0020] The present disclosure relates to extending the capabilities of an Al agent.
[0021] The present disclosure is defined by the scope of the independent claims. The dependent claims provide advantageous embodiments of the present disclosure.
[0022] The present disclosure provides in a first aspect a system for enhancing a first Al agent’s capabilities, the system comprising: a first Al agent configured to receive a query from a user and generate a first output;
[0023] a capability assessment module configured to determine whether the first Al agent can provide a sufficient solution to the query,
[0024] wherein the first Al agent is configured to
[0025] transmit information regarding the query to a second Al agent upon determining the first Al agent cannot provide a sufficient solution to the query;
[0026] a second Al agent configured to:
[0027] receive at least a portion of the query and / or relevant data from the first Al agent upon determining that the first Al agent is incapable of providing the sufficient solution; and
[0028] generate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge;
[0029] an update integration module configured to incorporate the one or more updates into the first Al agent,
[0030] wherein the second Al agent does not directly provide a response to the user but instead enables the first Al agent to address the query or future queries by virtue of the integrated updates.
[0031] This system addresses the limitations of existing automated prompt-generation solutions by providing a mechanism through which a second Al agent can produce and refine generalized instructions or “prompts” that extend the functional scope of afirst Al agent, rather than merely handling a single query in isolation. Conventional methods often generate query- specific prompts that cannot be reused, thereby failing to truly enhance an Al agent’s capabilities overtime. By contrast, the present system enables the second Al agent to create universal instructions tailored for a wide range of potential queries or tasks. These instructions are then integrated into the first Al agent’s instruction library, planner, tools, or knowledge base, allowing the first Al agent to solve new classes of queries without retraining or extensive augmentation of its underlying ML model. Moreover, because the second Al agent’s process can account for the first Al agent’s specific domain requirements and operational constraints, this approach supports various specialized Al agents while reducing reliance on expensive, human-crafted prompts.
[0032] According to an implementation of the first aspect, the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’ s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located.
[0033] Separating the first and second Al agents onto distinct network levels enables a flexible and distributed deployment. It accommodates scenarios where the first Al agent operates on lightweight hardware or an edge device, while the second Al agent’s more robust resources reside on a higher-tier or different network infrastructure. This arrangement also helps manage bandwidth and computational demands efficiently by letting each Al agent function in an environment suited to its role.
[0034] According to an implementation of the first aspect, the one or more updates are integrated during periods of low usage or scheduled downtime.
[0035] By scheduling the integration of new updates into the first Al agent during periods of reduced traffic, the system minimizes disruption to ongoing processes. This approach not only ensures uninterrupted service for users during peak periods but also optimizes resource usage, as it prevents high-demand tasks from competing with the more resource-intensive update processes.
[0036] According to an implementation of the first aspect, the first Al agent is more specialized than the second Al agent.
[0037] When the first Al agent has a specialized, domain-specific focus and the second Al agent is broader or more general-purpose, the user benefits from expert-level responses to routine queries. At the same time, the first Al agent can tap into the second Al agent’s wider range of knowledge or functionality for more complex tasks. This tiered structure combines specialized accuracy with the ability to handle edge cases or specialized knowledge gaps.
[0038] According to an implementation of the first aspect, the first Al agent’s capabilities are selectively updated based on at least one factor selected from:
[0039] a geographic location of the first Al agent,
[0040] a predefined first Al agent specialization or profile,
[0041] a time-based schedule.
[0042] Targeted updates enable the system to refine the first Al agent’s performance according to factors such as geographic location, time-based schedules, or specialized user profiles. This approach avoids a one-size-fits-all update strategy and makes the system highly adaptable to different environments, regulations, or use cases, ensuring that changes are both relevant and efficient.
[0043] According to an implementation of the first aspect, the first Al agent stores or logs indications of unsolvable queries in a buffer and transmits said queries in bulk to the second Al agent during off-peak hours to reduce resource usage.By collecting unsolvable or complex queries in a buffer, the first Al agent can batch them for transmission to the second Al agent during non-peak hours. This method conserves bandwidth and processing power, and it prevents downtime during critical usage periods.
[0044] According to an implementation of the first aspect, the system further comprises:
[0045] a request forming module configured to prepare the information regarding the query based on at least a portion of the query and / or data pertaining to the query.
[0046] A request forming module refines and packages query data before sending it to the second Al agent. This step standardizes communication, ensuring consistency in how queries are interpreted and processed. It can also incorporate privacy safeguards or data-minimization strategies to protect user information while still enabling the second Al agent to generate accurate updates.
[0047] According to an implementation of the first aspect, the second Al agent is further configured to generate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent; and
[0048] if the new task instructions require it, providing at least one of new tools and knowledge for the first Al agent.
[0049] By generating universal instructions for the first Al agent, the second Al agent provides improvements in a structured, systematic manner. This centralized instruction generation ensures that updates to the first Al agent’s instructions, tools, or knowledge base follow a unified format and can be implemented consistently. The user is supplied with the improved capabilities in the first Al agent, while the underlying updates may not require direct user interaction.
[0050] According to an implementation of the first aspect, the universal instruction generator is configured to:
[0051] create a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent;
[0052] input the single task instruction and any task example into a model of the first Al agent;
[0053] refine the single task instruction based on solutions provided by the model of the first Al agent; and
[0054] continue until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.
[0055] In creating a single task instruction, the second Al agent iteratively refines the instructions needed for solving a generalized class of tasks. This process involves comparing trial solutions from the first Al agent’ s model against a success threshold, thus guaranteeing that each instruction meets a predetermined level of accuracy. The user ultimately benefits from well-tested improvements without experiencing trial-and-error in real time.
[0056] According to an implementation of the first aspect, the single task instruction is a structured description that can be parsed by the first Al agent and wherein the single task instruction contains a task description, resolving procedure and desired output and wherein the single task instruction is further tailored to the first Al agent according to its specialization.
[0057] The final single task instruction may be delivered in a form that the first Al agent can readily parse and execute, incorporating a task description, a procedure for resolving the query, and an indication of the desired output. By tailoring these instructions to the specific domain or specialization of the first Al agent, the system can ensure that updates are both targeted and efficiently integrated, thereby enhancing the first Al agent’s performance.The present disclosure provides in a second aspect a method of enhancing a first Al agent’s capabilities, the method comprising: receiving a query at a first Al agent from a user;
[0058] generating a first output or confidence metric via the first Al agent;
[0059] determining whether the first Al agent can provide a sufficient solution based on the first output or confidence metric; transmitting information regarding the query to a second Al agent upon determining the first Al agent is insufficient; obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature; and
[0060] integrating the one or more updates into the first Al agent, wherein the second Al agent does not provide a direct response to the user.
[0061] This method represents process-based approach to the problem at hand. It delineates how the first Al agent receives a query, assesses its solvability, consults the second Al agent for additional support, and integrates updates that expand the first Al agent’s ability to address new or difficult tasks. Such a method-focused perspective offers a clear procedural blueprint for implementation, covering diverse hardware setups and network architectures.
[0062] According to an implementation of the second aspect, the information regarding the query comprises a request and relevant data obtained through a request forming module, based on at least a portion of the query and / or data pertaining to the query and wherein the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located.
[0063] Using a request forming module in conjunction with or separate to network-level separation ensures consistent, standardized communication between the first and second Al agents, regardless of their respective locations. This approach clarifies the steps needed to prepare a query for the second Al agent while also allowing each agent to reside on infrastructure that best matches its performance and resource requirements.
[0064] According to an implementation of the second aspect, the one or more updates are integrated during periods of low usage or scheduled downtime and wherein the first Al agent is more specialized than the second Al agent.
[0065] Scheduling updates during times of low traffic, coupled with having the second Al agent complement the first Al agent’s specialized domain knowledge, allows the system to run more efficiently. It avoids performance bottlenecks during high-usage intervals and ensures that specialized tasks remain accurate thanks to the methodical support of a broader, more general second Al agent.
[0066] According to an implementation of the second aspect, before obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature the method comprises:
[0067] generating one or more updates applicable to a first Al agent’ s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent; and if the new task instructions require it, providing at least one of new tools and knowledge for the first Al agent.
[0068] By generating updates in the form of new task instructions, the method ensures that improvements are made systematically before reaching the first Al agent. This step reduces trial-and-error for the user, as the second Al agent refines the instructions to meet predefined success thresholds, producing updates that are well-tested and ready for immediate application.According to an implementation of the second aspect, the method further comprises providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent comprises:
[0069] creating a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent;
[0070] inputting the single task instruction and any task example into a model of the first Al agent;
[0071] refining the single task instruction based on solutions provided by the model of the first Al agent; and
[0072] continuing until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.
[0073] Through iterative refinement, each instruction is tested by the first Al agent’s model and compared against a threshold of acceptable performance. This cycle continues until the instructions achieve a level of quality sufficient for real- world use. As a result, users benefit from increasingly sophisticated solutions generated by the first Al agent, while the refining process remains fully contained within the second Al agent’s operational pipeline.
[0074] BRIEF DESCRIPTION OF DRAWINGS
[0075] In the following embodiments of the present disclosure are described in more detail with reference to the attached figures and drawings, in which:
[0076] Fig. 1 shows schemes for sending common and new queries to a specialized Al agent. Specifically, on the left a setting with end users, e.g. autonomous vehicles, robots, mobile phones, etc. sending queries to a specialized Al agent in a mobile network, and receiving solutions is considered. On the right, a problem of extending the capabilities of an Al agent in order to solve new types of queries from end users is considered.
[0077] Fig. 2 shows on the left a scenario when a received query and data are within a first Al Agent or Agent A’s specialization domain and on the right a received query and / or data is outside the specialization domain, and Agent A must extend its capabilities in order to solve the user’s query and similar queries in the future.
[0078] Fig. 3 illustrates a procedure for generating a good task instruction by a second Al agent or Agent B.
[0079] Fig. 4 is an Illustration of a next-generation mobile network with multiple specialized Al agents at Radio Access Network, RAN, and a universal agent at Core Network, CN, helping the specialized agents to evolve.
[0080] Fig. 5 is a flow diagram of a method presented herein.
[0081] DETAILED DESCRIPTION
[0082] In the description of embodiments of this application, "and / or" describes an association relationship between associated objects and indicates that three relationships may exist. For example, A and / or B may indicate the following three cases: Only A exists, both A and B exist, and only B exists. In this application, "at least one" means one or more, and "a plurality of' means two or more. In addition, it should be understood that in description of this application, terms such as "first", "second", and "third" are merely used for distinguishing and description, but should not be understood as indicating or implying relative importance, or should not be understood as indicating or implying a sequence.
[0083] As used herein, the term “Al agent” is understood to mean any entity - embodied in hardware, software, firmware, or any combination thereof - that is capable of receiving data or user input, making decisions or inferences based on one or more computational models, and producing an output that addresses the given query or scenario. For example, an Al agent may include:
[0084] One or more machine learning, ML, models, such as deep neural networks, classical algorithms, or specialized modules for speech recognition, text processing, or data classification.Rule-based expert systems or any other logic-based decision modules capable of outputting meaningful responses to user queries.
[0085] Virtual environments, e.g. containerized or serverless architectures, that run ML code and provide inference or training capabilities.
[0086] Al accelerator chips, e.g., GPUs, TPUs, NPUs, FPGAs, ASICs, deployed on local edge devices or embedded systems that facilitate faster data processing and ML inference.
[0087] Dedicated modules or circuits such as system-on-chip, SoC, components with built-in Al instructions or integrated neural processing units.
[0088] Smart loT devices or robotics platforms with sensors, actuators, memory, and processor resources for on-device machine learning or real-time analytics.
[0089] Combinations of local and remote resources, where certain Al functionalities are executed on a user device, e.g. smartphone, tablet, wearable device, and others run on a remote server or cloud platform.
[0090] Edge computing paradigms that split inference tasks between local processing units and large-scale data centers.
[0091] In short, an Al agent need not be purely software-based and can optionally incorporate specific hardware features or purpose-built circuits to carry out Al-related tasks.
[0092] In the context of high-speed mobile networks such as 5G and 6G, data communication and service provisioning are generally structured into distinct layers or segments, including but not limited to:
[0093] The Radio Access Network, RAN, is responsible for establishing and managing the radio, wireless, link between end-user devices, e.g. smartphones, loT terminals, and the cellular network.
[0094] It involves base stations, remote radio heads, antennas, and other infrastructure components that handle signal modulation, encoding, and decoding.
[0095] In 5G and 6G, the RAN may incorporate massive MIMO, Multiple Input, Multiple Output, technology, beamforming, and ultra-dense network topologies to deliver high throughput and low latency.
[0096] The core network, CN manages higher-level services such as user authentication, data routing, session management, and policy enforcement.
[0097] In modem architectures, e.g. 5G standalone cores, the network adopts a service-based architecture, SBA, featuring microservices and network slicing for different use-case scenarios, e.g., enhanced mobile broadband, ultra-reliable low-latency communication, massive machine-type communication.
[0098] The CN is typically more centralized or regionally distributed compared to the RAN and is designed to interconnect with other networks, including the internet or private enterprise networks.
[0099] Depending on deployment scenarios, other layers or infrastructure components, e.g. edge computing nodes, local cloudlets, may sit between the RAN and the CN to offload processing, including Al inference tasks.
[0100] Consequently, an Al agent in the present disclosure could operate at the device level within the RAN, e.g., at the user’s smartphone or loT endpoint, within the core network, e.g. as part of a server-side cloud or cluster that handles large-scale ML tasks, or in a hybrid arrangement that leverages edge and central resources to optimize performance, latency, and resourceusage. By acknowledging these various network layers and hardware configurations, the invention remains flexible and adaptable to existing and future telecommunication standards, including but not limited to 5G and 6G environments.
[0101] A task instruction, also denoted as an instruction prompt, is a structured description of the task that is used by an Al agent when processing a user’s query. The instruction helps to better understand the problem to be solved by the agent, it decomposes difficult tasks into manageable steps, and it guides the agent’s specialized Al model towards desired outcomes. Furthermore, a single task instruction could be used to solve not just a single specific query, but any query from a given semantic class, e.g., any obstacle avoidance query. Thereby, instruction tasks help to strike a good trade-off between the agent’s generalization and specialization capabilities.
[0102] Adding a new instruction prompt to an Al agent’s library enables extending the agent’s capabilities in multiple ways. Firstly, an instruction can describe how a specialized agent can use new tools and programs. Secondly, an instruction can provide a new resolving strategy based on already available tools and knowledge, allowing the agent to perform more sophisticated tasks. Finally, an instruction can introduce new concepts based on concepts already known by the agent, expanding the agent’s domain of specialization. For example, a car-centric agent can learn concepts of a bicycle using already known car-specific vocabulary, making the agent vehicle-centric instead of only car-centric. The above reasons clearly highlight that instructions should be adapted to a particular agent, and that an instruction which works well for one agent may fail completely for another.
[0103] Task instructions for Al agents may be written in simple text format, offering valuable interpretability for human readers. Any other format that can be parsed by the agent is also possible, including JSON documents, scripts in some programming language, and even images or videos. Irrespectively of the format, each instruction preferably keeps a clear structure, and include information such as task description, a resolving procedure, a list of required tools and knowledge, constraints, and a desired output. An example instruction structure could be as follows:
[0104] a. Introduction.
[0105] b. Task Description.
[0106] c. Examples with explanation.
[0107] d. Some tricks to solve the task.
[0108] e. Background Knowledge.
[0109] f. Rules or Constraints.
[0110] g. Challenges or comer cases.
[0111] h. Desired output format.
[0112] According to the above structure, an example of a task instruction could be:
[0113] a. “Description of how {user} can pass by {obstacle}”.
[0114] b. “The procedure is as follows:
[0115] a. Identify {user} type (e.g., a pedestrian, vehicle, bike) using the received metadata,
[0116] b. Identify {obstacle} (e.g. a vehicle, a tree), using the YOLO object detection algorithm and blueprints stored in the knowledge database,
[0117] c. Find a safe way around {obstacle} by using ...”
[0118] c. “Similar scenarios include ...”
[0119] d. “Take perspective of {user} and think step-by-step. ...”
[0120] e. Etc.The way how a task instruction is used by an agent depends on implementation. For example, given a query, the agent can select a most suitable instruction and append it to the query in a similar way like prompts. Another possibility is to use the information from the query to adapt the instruction, and execute only the instruction. Finally, the Al agent could be equipped with two coupled Al models which take as input the query and the instruction separately, and combine the models’ outputs together.
[0121] Fig. 2 shows a system for enhancing a first Al agent’s, or Agent A’s, capabilities, the system comprising a first Al agent configured to receive a query from a user and generate a first output, a capability assessment module configured to determine whether the first Al agent can provide a sufficient solution to the query, wherein the first Al agent is configured to transmit information regarding the query to a second Al agent upon determining the first Al agent cannot provide a sufficient solution to the query, a second Al agent configured to receive at least a portion of the query and / or relevant data from the first Al agent upon determining that the first Al agent is incapable of providing the sufficient solution and generate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge, and an update integration module configured to incorporate the one or more updates into the first Al agent, wherein the second Al agent does not directly provide a response to the user but instead enables the first Al agent to address the query or future queries by virtue of the integrated updates.
[0122] Fig. 2 shows on the left a scenario when a received query and data are within a first Al agent’s specialization domain and on the right a received query and / or data is outside the specialization domain, and the first Al agent must extend its capabilities in order to solve the user’s query and similar queries in the future. A capability assessment module, which may be part of a query interpreter, is configured to determine whether the first Al agent can provide a sufficient solution to the query. Optionally, the first Al agent may determine at any step, i.e. when employing an instruction library and planner and / or when employing tools and knowledge through a query executor, that it is incapable of solving a request. A request may be information, instruction data, or a task specification submitted to the Al agent for processing or evaluation, typically comprising one or more query parameters, contextual details, or any other relevant data needed by the Al agent to produce an appropriate response. The capability assessment module may in such a situation be a standalone module accessed by the query interpreter and / or the query executor or part of each of the query interpreter and the query executor. In the situation shown on the left side of Fig. 2, the first Al agent, or Agent A as depicted in Fig. 2, receives a query from a user, determines that it is capable of providing an output and is thus configured to and generate a first output. When the first Al agent determines that it cannot provide a sufficient solution to the query, it is configured to transmit information regarding the query to a second Al agent, or Agent B as depicted in Fig. 2 on the right. The second Al agent is configured to receive at least a portion of the query and / or relevant data from the first Al agent upon determining that the first Al agent is incapable of providing the sufficient solution. Said portion may be determined by a specific request forming module in the first Al agent. The second Al agent is further configured to generate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge. An update integration module is configured to incorporate the one or more updates into the first Al agent. The second Al agent does not directly provide a response to the user but instead enables the first Al agent to address the query or future queries by virtue of the integrated updates. This arrangement ensures that users only interact with the first Al agent while benefiting from improvements introduced through the second Al agent. The second Al may optionally comprise, as depicted, a request interpreter receiving the at least a portion of the query and / or relevant data from the first Al agent, a universal instruction generator and a separate tools and knowledge module, wherein the universal instruction generator can be tasked with providing new task instruction to the instruction library and planner of the first Al agent and the separate tools and knowledge module can be tasked with providing new tools and knowledge to be added to the tools and knowledge module of the first Al agent.Summarizing still with regard to Fig. 2, when an end user sends a query to the first Al agent or Agent A, two scenarios are possible:
[0123] • The received query and data are within the Agent A’s specialization domain, and could be solved using a standard procedure,
[0124] • The received query and / or data is outside the specialization domain, and Agent A must extend its capabilities in order to solve the query and similar queries in the future.
[0125] A query describes a user’s intent, and may include a desired outcome, requirements, and constraints. Furthermore, a query can be accompanied by additional data including sensing data and context metadata, e.g., data from a vehicle’s radar, technical information about the vehicle itself, etc.
[0126] As mentioned, a task instruction, also denoted as an instruction prompt, is a structured description of the task to be solved including a general resolving procedure, a list of required tools and knowledge, constraints, and a desired output. It may be a simple text format, but other formats are also possible, e.g., a JSON document, a script in some programming language, etc.
[0127] a) If received query and data are within the specialization domain (standard procedure) of the first Al agent or Agent A as depicted, an example procedure would be
[0128] i. Agent A interprets the query in the view of received data and meta-data, but also in the view of information available at the network, e.g., network traffic, meta-data, network sensors, etc.),
[0129] ii. Agent A creates a plan for solving the query, including identifying all types of information needed to solve the query, e.g., local traffic information, city maps, vehicle blueprints, etc., required tools and programs to be executed, e.g., a calculator, an Al-based object detection algorithm, a shortest-path navigation algorithm, a web-browser, etc., and a procedure to follow in order to obtain the desired result.
[0130] Planning the query execution is done using the Agent A’s specialized Al model. Usually, this step requires advanced reasoning capabilities, and it is resource- and time-consuming. To simplify this step, Agent A can use pre-defined task instructions stored in a built-in instruction library. Instead of planning the query execution each time from the very beginning, the agent can select a suitable task instruction from its library, adapt the selected instruction to the received query, and use the instruction as a prompt.
[0131] iii. Agent A loads necessary knowledge from its knowledge database and tools from its toolbox.
[0132] iv. Agent A uses the loaded knowledge and tools to execute the procedure on the received query, and obtains a solution. This step does not require advanced reasoning capabilities, as opposed to planning.
[0133] v. Finally, Agent A sends back a solution to the end user.
[0134] b) If the received query and data are outside the specialization domain of the first Al agent or Agent A an example procedure would be
[0135] i. Agent A interprets the query in the view of received data and meta-data, but also in the view of information available at the network, e.g., network traffic, meta-data, sensors, etc. The agent detects that it cannot reliably solve the query due to insufficient knowledge, missing tools, and / or lack of resolving procedures.
[0136] ii. Agent A forms a request for solving the query. The request includes the end user’s query and data, but also other information from Agent A like own query and data.
[0137] iii. Agent A sends a message to Agent B with the ID of Agent A, the request, additional data, and a timestamp:
[0138] [ID Agent A, request, data, timestamp].iv. Agent B receives the message and interprets the request.
[0139] v. Agent B uses its foundation model as a Universal Instruction Generator, and generates a new instruction which describes a general procedure for solving the user’s query, and which is adapted for Agent A.
[0140] vi. Agent B, based on the generated instructions, loads required knowledge from its own large knowledge database, and tools from its own large toolbox.
[0141] vii. Agent B sends back a message to Agent A with the ID of Agent B, the instruction, tools, additional data, and a timestamp: [ID Agent B, instruction, tools, data, timestamp].
[0142] viii. Agent A receives the messages and loads the instruction to its instruction library, knowledge data to its knowledge database, and tools to its toolbox. The capabilities of Agent A are now augmented, and the agent can now solve the end user’s query and similar queries in the future.
[0143] Optionally, the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located. An advantage of such an arrangement is that the more lightweight or specialized first Al agent can operate at an edge level, while the second Al agent runs at a core level with broader coverage or additional resources.
[0144] Optionally, the one or more updates are integrated during periods of low usage or scheduled downtime, which minimizes impacts on system performance.
[0145] Optionally, the first Al agent is more specialized than the second Al agent, which allows a narrow domain expert to tap into broader updates when faced with out-of-scope queries.
[0146] Optionally, the first Al agent’s capabilities are selectively updated based on at least one factor selected from a geographic location of the first Al agent, a predefined first Al agent specialization or profile and a time-based schedule. This selective update approach can further optimize resource use by ensuring that only relevant improvements are applied to a specific first Al agent.
[0147] Optionally, the first Al agent stores or logs indications of unsolvable queries in a buffer and transmits said queries in bulk to the second Al agent during off-peak hours to reduce resource usage. This buffering improves efficiency by reducing traffic load, especially with regard to wireless networks in high density areas.
[0148] Optionally, the system further comprises a request forming module configured to prepare the information regarding the query based on at least a portion of the query and / or data pertaining to the query. This request forming module can structure the data in a standardized format, aiding the second Al agent in swiftly generating the one or more updates for the first Al agent.
[0149] Fig. 3 illustrates an optional system for generating a good task instruction by the second Al agent, or Agent B as depicted. Optionally, the second Al agent is further configured to generate one or more updates applicable to a first Al agent’ s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent; and, if the new task instructions require it, provide at least one of new tools and knowledge for the first Al agent.
[0150] Optionally, the universal instruction generator is configured to create a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent, input the single task instruction and any task example into a model of the first Al agent, refine the single task instruction based on solutions provided by the model ofthe first Al agent, and continue until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.
[0151] Further optionally, the single task instruction is a structured description that can be parsed by the first Al agent and wherein the single task instruction contains a task description, resolving procedure and desired output and wherein the single task instruction is further tailored to the first Al agent according to its specialization.
[0152] Given a query from the first Al agent or Agent A as depicted in Fig. 3, the main objective of the instruction generator is to create a task instruction that is generic enough to cover a range of similar queries, and is adapted to the knowledge and characteristics of Agent A. To achieve this goal, this embodiment may utilize a novel concept based on Learning from Feedback and its improved version called Direct Preference Optimization. The concept of the system may be to introduce a feedback loop between the Universal Instruction Generator and the target agent or its local copy. The feedback loop optimally enables the generator to produce an optimal task instruction via a dialogue with the target agent.
[0153] Fig. 3 illustrates a diagram of a system and procedure for producing an optimal task instruction by the instruction generator. The generation of the instruction can be done fully inside the second Al agent or Agent B, i.e., via internal dialogue, or by exchanging feedback information with another agent. A sample system and procedure is listed below:
[0154] a) Upon receiving the request from Agent A, Agent B creates multiple similar versions of the query, e.g., with different query formulation, details, images, etc. This can be done using a generative model or by searching in a large database. The result is a set of task examples which generalize the initial query.
[0155] b) The Universal Instruction Generator, which may be a foundation model, at Agent B creates a single task instruction for solving any task example from the example dataset.
[0156] c) The instruction and the examples are given to a clone of Agent A.
[0157] d) Agent A tries to solve all examples using the instruction, and outputs result, namely successful or failed outputs. e) The results are forwarded back to the Universal Instruction Generator which refines the instruction.
[0158] f) The procedure continues until a good task instruction task is found, i.e., the target agent is able to solve all or nearly all queries from the example dataset.
[0159] The result of the procedure is a task instruction which is general, i.e. it describes a solving procedure for a class of queries, and, at the same time, is adapted to the target agent since the target agent provided its feedback during instruction generation.
[0160] Fig. 4 shows a possible implementation of Agentic Al in a next-generation mobile network with the present invention for extending capabilities of specialized agents. The presented framework consists of three layers: a level of mobile end users with User Equipment, UE, a level of Radio Access Network, RAN, with specialized first Al agents, and a level of Core Network, CN, with a universal second Al agent equipped with a foundation model acting as a Universal Instruction Generator.
[0161] In this scenario, intelligent end users such as cars, cellphones or other devices perform autonomously or semi-autonomously various complex task such as autonomous driving, smart manufacturing, robotic collaboration. Whenever the users encounter challenges they cannot solve by themselves, they send queries to the closest access point equipped with a local specialized Al agent. The specialized agent uses own knowledge and resources to solve the query and send back a response to end users. However, when the specialized agent is incapable of solving the query, it sends a request to the universal agent at the Core Network in order to receive a suitable task instruction that expands the agent’s capabilities.In the following, methods pertaining to enhancing a first Al agent’s capabilities are described. General definitions and procedures already described with regard to the system above are not repeated to allow for a concise and clear description. The general procedural mechanism as embodied in and described for Figures 2, 3 and 4, however, still apply.
[0162] Fig. 5 is a flow diagram of a method presented herein. The method of enhancing a first Al agent’s capabilities comprises receiving a query at a first Al agent from a user SI, generating a first output or confidence metric via the first Al agent S2, determining whether the first Al agent can provide a sufficient solution based on the first output or confidence metric S3, transmitting information regarding the query to a second Al agent upon determining the first Al agent is insufficient S4, obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature S5 and integrating the one or more updates into the first Al agent, wherein the second Al agent does not provide a direct response to the user S6.
[0163] A confidence metric is a quantitative or qualitative indicator generated by a computational model, rule-based mechanism, or statistical process to measure how likely the output of a given Al system is correct or sufficient. This metric can take various forms, such as a probability score, a numerical ranking, a threshold-based classification, or any other evaluative indicator that conveys the Al system’s assessed reliability or certainty in its result.
[0164] An insufficient solution is a solution that fails to satisfy one or more predetermined criteria for completeness or correctness, such as falling below a required confidence threshold or lacking necessary domain coverage, thereby indicating that the Al agent’s output is not adequately reliable for the query at hand.
[0165] As to the general process and the definition of terms like query, reference is made to Fig. 2 and the description pertaining to the system above.
[0166] Optionally, the information regarding the query comprises a request and relevant data obtained through a request forming module, based on at least a portion of the query and / or data pertaining to the query and wherein the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located.
[0167] Optionally, the one or more updates are integrated during periods of low usage or scheduled downtime and wherein the first Al agent is more specialized than the second Al agent.
[0168] Optionally, before obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature the method comprises generating one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent and, if the new task instructions require it, providing at least one of new tools and knowledge for the first Al agent.
[0169] Optionally, providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent comprises creating a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent, inputting the single task instruction and any task example into a model of the first Al agent, refining the single task instruction based on solutions provided by the model of the first Al agent and continuing until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.As to further details and sample procedures regarding this process, see above with regard to Fig. 3.
[0170] The foregoing description merely provides examples of specific implementations of this application but are not intended to limit the protection scope of this application. Any variation or replacement readily figured out by a person skilled in the art within the technical scope disclosed in this application shall fall within the protection scope of this application. Therefore, the protection scope of this application shall be subject to the protection scope of the claims.
Claims
CLAIMS1. A system for enhancing a first Al agent’s capabilities, the system comprising:a first Al agent configured to receive a query from a user and generate a first output;a capability assessment module configured to determine whether the first Al agent can provide a sufficient solution to the query,wherein the first Al agent is configured totransmit information regarding the query to a second Al agent upon determining the first Al agent cannot provide a sufficient solution to the query;a second Al agent configured to:receive at least a portion of the query and / or relevant data from the first Al agent upon determining that the first Al agent is incapable of providing the sufficient solution; andgenerate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge;an update integration module configured to incorporate the one or more updates into the first Al agent,wherein the second Al agent does not directly provide a response to the user but instead enables the first Al agent to address the query or future queries by virtue of the integrated updates.
2. The system of claim 1, wherein the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located.
3. The system of claims 1 or 2, wherein the one or more updates are integrated during periods of low usage or scheduled downtime.
4. The system of any of claims 1 to 3, wherein the first Al agent is more specialized than the second Al agent.
5. The system of any of claims 1 to 4, wherein the first Al agent’s capabilities are selectively updated based on at least one factor selected from:a geographic location of the first Al agent,a predefined first Al agent specialization or profile,a time-based schedule.
6. The system of any of claims 1 to 5, wherein the first Al agent stores or logs indications of unsolvable queries in a buffer and transmits said queries in bulk to the second Al agent during off-peak hours to reduce resource usage.
7. The system of any of claims 1 to 6, wherein the system further comprises:a request forming module configured to prepare the information regarding the query based on at least a portion of the query and / or data pertaining to the query.
8. The system of any of claims 1 to 7, wherein the second Al agent is further configured to generate one or more updates applicable to a first Al agent’s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent; andif the new task instructions require it, provide at least one of new tools and knowledge for the first Al agent.
9. The system of claim 8, wherein the universal instruction generator is configured to:create a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent;input the single task instruction and any task example into a model of the first Al agent;refine the single task instruction based on solutions provided by the model of the first Al agent; andcontinue until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.
10. The system of claim 9, wherein the single task instruction is a structured description that can be parsed by the first Al agent and wherein the single task instruction contains a task description, resolving procedure and desired output and wherein the single task instruction is further tailored to the first Al agent according to its specialization.
11. A method of enhancing a first Al agent’s capabilities, the method comprising:receiving a query at a first Al agent from a user (SI);generating a first output or confidence metric via the first Al agent (S2);determining whether the first Al agent can provide a sufficient solution based on the first output or confidence metric (S3); transmitting information regarding the query to a second Al agent upon determining the first Al agent is insufficient (S4); obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature (S5); andintegrating the one or more updates into the first Al agent, wherein the second Al agent does not provide a direct response to the user (S6).
12. The method of claim 11, wherein the information regarding the query comprises a request and relevant data obtained through a request forming module, based on at least a portion of the query and / or data pertaining to the query and wherein the first Al agent and the second Al agent are accessible via a network service, such that the first Al agent is not necessarily located on a user’s local device and wherein the first Al agent is located on a first network level different from a second network level on which the second Al agent is located.
13. The method of claims 11 or 12, wherein the one or more updates are integrated during periods of low usage or scheduled downtime and wherein the first Al agent is more specialized than the second Al agent.
14. The method of any of claims 11 to 13, wherein before obtaining from the second Al agent one or more updates that enable the first Al agent to address the query or future queries of a similar nature the method comprises:generating one or more updates applicable to a first Al agent’ s instruction library, planner, tools and / or knowledge by providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent; and if the new task instructions require it, providing at least one of new tools and knowledge for the first Al agent.
15. The method of any of claim 14, wherein providing new task instructions for the first Al agent through a universal instruction generator in the second Al agent comprises:creating a single task instruction for solving any task example from an example dataset generalized from the initial query by the second Al agent;inputting the single task instruction and any task example into a model of the first Al agent;refining the single task instruction based on solutions provided by the model of the first Al agent; andcontinuing until a good single task instruction is found as defined by a predetermined success threshold for the solutions provided by the model of the first Al agent.