Abstract: State-of-the-art Integration Platform as a Service (iPaaS) platforms do not enable users to utilize their own artificial intelligence (AI) models (e.g., large language models or other generative model). Disclosed embodiments enable users to bring your own model (BYOM) within an integration platform. In particular, users may build or import their own AI model, and integrate that AI model into their integration processes in the same manner as any other integration step. This may involve dragging a shape, representing the AI model, from a virtual palette onto a virtual canvas, and configuring the corresponding AI step. During execution, the resulting integration process may automatically and dynamically, in real time, receive input data and retrieve contextual metadata to enhance the input data with context. The combination of input data and contextual metadata may be provided to the AI step to perform the desired task with improved accuracy.

Abstract: Current agentic systems are unable to balance the costs of artificial intelligence (AI) agents with performance. Disclosed embodiments introduce a scoring AI agent, which scores deviations between the actual performance of a performing AI agent and the expected performance of the performing AI agent. When this deviation becomes substantial, the scoring AI agent may modify parameter(s) in an adaptive governance policy that governs operation of the performing AI agent. In this manner, an AI agent can be automatically throttled down and/or up, as real-time conditions evolve.

Abstract: Conventional testing of artificial intelligence (AI) agents is time-consuming and insufficient. Accordingly, embodiments provide an automated testing and validation framework for the tools utilized by AI agents. The framework may leverage artificial intelligence, including a classifier for classifying tools, a generative model that creates test cases based on the tool classification, an optimization model that maximizes test coverage while minimizing computation cost, an edge-detection model that identifies edge test cases, and/or an execution engine that manages execution of the test suite. The framework may be integrated into a continuous integration and continuous deployment (CI/CI) pipeline.

Abstract: Existing platforms for developing artificial intelligence (AI) agents rely on static testing or limited real-world scenarios to evaluate agentic behavior. Accordingly, a simulator is disclosed that enables the simulation of an AI agent with adaptive test scenarios. During the simulation, a visualization interface may display a trace of the behavioral flow of the AI agent in real time. The simulator may also monitor and display one or more performance metrics and provide indications of the guardrail compliance of the AI agent in real time. A user may also pause, rewind, or modify the simulation in real time, to make immediate and real-time adjustments to the configuration of the AI agent.

Abstract: There is currently no tool for predicting the cost of performing an inference by an artificial intelligence (AI) agent. As a result, AI agents may easily end up exceeding economic, computational, energy, and/or ecological budgets. Disclosed embodiments utilize one or more AI agents to identify similar historical inputs to a prospective input, on which inference is to be performed, and predict an economic, computational, energy, and/or ecological cost of performing inference on the prospective input. If the cost would exceed a budget, the inference may be automatically blocked, at least temporarily, to prevent the budget from being exceeded.

Abstract: A federated data system enables multiple, independent, decentralized, and autonomous data sources to be queried as if they were a single system. State-of-the art federated data systems suffer from excessive data movement, static execution strategies, and inefficient distributed joins. Disclosed embodiments introduce a multi-agent framework into the federation layer of the federated data system. This agentic framework may comprise a supervisor artificial intelligence (AI) agent that divides queries, from a query engine, into subqueries and utilizes other AI agents to perform sub-tasks, such as rewriting one or more the of the subqueries, determining an execution strategy, executing the subqueries according to the execution strategy, and learning from past query executions. The supervisor AI agent may perform in-memory joins of the results of the subqueries, and return a final query result to the query engine.

Abstract: There is no unified, comprehensive system for the governance of artificial intelligence (AI) agents. Accordingly, disclosed embodiments register a plurality of AI agents, potentially from disparate sources, into a centralized registry. Prior to activation, each AI agent is associated with a governance policy. After activation, each AI agent is monitored in real time, during execution, to determine whether or not the AI agent remains compliant with the associated governance policy. When an AI agent becomes non-compliant with the associated governance policy, corrective action may be taken, and the AI agent may be labeled as untrusted. The corrective action may comprise alerting a user, modifying the AI agent, adjusting an amount of each of one or more computational resources that is allocated to the AI agent, modifying an access of the AI agent to one or more systems, adjusting a communication control of the AI agent, and/or the like.

Abstract: Conventionally, artificial intelligence (AI) agents must be manually registered with platforms, which results in a lack of scalability, as the number of AI agents continues to surge into the millions and hundreds of millions, a lack of compatibility and standardization between agent specifications, and siloes of AI agents that cannot be combined into a unified registry. Accordingly, disclosed embodiments provide a AI-powered registration service that is capable of automatically registering AI agents, regardless of input format and regardless of the framework in which those AI agents are defined, using a standard agent schema, into a unified and centralized registry. In turn, this unified registry improves searching and discovery of AI agents.

Abstract: There is currently no protocol that enables artificial intelligence (AI) agents to autonomously select provider entities in a scalable, standardized, transparent manner, based on real-time conditions. In disclosed embodiments, provider entities autonomously register with a distributed ledger. Consumer AI agents may, when needing a new provider entity, autonomously query the distributed ledger and select a new provider entity. When selecting the new provider entity, a consumer AI agent may execute a negotiation protocol to obtain an optimal service agreement. In this manner, provider entities may autonomously onboard themselves, and AI agents may autonomously evolve, over time, to acquire new capabilities, by selecting new provider entities from an ever expanding universe of autonomously onboarded provider entities.

Abstract: As artificial intelligence (AI) agents become more prevalent, it has become important to measure their effectiveness. Disclosed embodiments enable autonomous, real-time evaluation of AI agents using a monitoring service and peer AI agents. In an embodiment, calls, by a performing AI agent, to models and tools, during a session, are made through respective gateways which collect session data. A monitoring service acquires the session data from the gateways, and invokes one or a plurality of monitoring AI agents to evaluate the performance of the performing AI agent based on the session data and one or more adaptable session parameters. The result of the evaluation(s) may be stored for analysis and development of the performing AI agent.

Abstract: Artificial intelligence (AI) agents are increasingly being used to automate tasks within integration environments. However, a problem occurs in hybrid environments in which the AI agent is hosted in a cloud-computing environment, whereas the tools, required by the AI agent, are hosted on an on-premise system of an organization. In this case, the AI agent may not know the organization-specific operations available through the on-premise tool—let alone, the requirements of those operations. Accordingly, embodiments enable an AI agent to automatically learn the specific operations, including the requirements (e.g., inputs and outputs) of those operations, during installation and/or execution, via a predefined discovery operation that is implemented by the application programming interface of every on-premise tool.

Abstract: Typically, in order to preserve their specializations, AI agents operate with isolated knowledge bases to learn skills independently from other AI agents. This results in a waste of computational resources and reduces the adaptability of the AI agents and overall system. Accordingly, disclosed embodiments enable AI agents to learn from each other, while retaining their specialized capabilities. A knowledge manager may identify knowledge elements, within a knowledge repository, that are compatible with a receiving AI agent and do not diminish the specialization of the receiving AI agent. An incremental transfer process may be executed, with performance monitoring and rollback capabilities, to incorporate the knowledge elements into the receiving AI agent, while ensuring that the knowledge transfer does not diminish the specialization of the receiving AI agent. A cross-agent collaboration interface is also disclosed for knowledge sharing and collaboration between AI agents.

Abstract: State-of-the-art tracing systems struggle with the analysis and visualization of the debugging data for artificial intelligence (AI) agents. In an embodiment, a trace engine obtains telemetry data during execution of an AI agent, and generates a hierarchical trace structure, comprising a decision trace structure representing a decision layer of the AI agent, an operation trace structure representing an operation layer of the AI agent, and an implementation trace structure representing an implementation layer of the AI agent. A visual debugging interface may query this hierarchical trace structure to generate one or more interactive visual elements for debugging of the AI agent.

Abstract: State-of-the-art platforms for the development of artificial intelligence (AI) agents typically require users to manually define and configure each component of an AI agent. This results in a high barrier of entry for new AI agents, a high likelihood of errors, long development cycles, and inefficiencies, especially for non-technical users. Accordingly, embodiments use artificial intelligence to generate a complete specification of an AI agent based on a user's intent, for example, as expressed in natural language. In an embodiment, the intent is converted into a structured intent, from which one or more tasks are determined. Next, one or more tools are identified for each task, and one or more guardrails are generated for the AI agent. The user may modify the AI-agent specification, as needed or desired, via an intuitive interface, before the AI agent is generated and deployed.

Abstract: As the number of AI agents in operation has increased, it has become difficult to analyze and search registries of AI agents. Disclosed embodiments enable a holistic view of the behaviors of AI agents, including governance and/or human-like behaviors, both individually and in relation to the behaviors of other AI agents. In particular, a profiling service may embed each of a plurality of behaviors of each AI agent into the same vector space by feeding heterogeneous data about the AI agent through an AI model, converting the output of the AI model into an embedding vector, and storing the embedding vector in a vector database in association with the AI agent. The behaviors that are embedded into the vector space may comprise governance behaviors, human-like behaviors, and/or any other category of behavior. This provides scalability and explainable searches and comparisons of AI agents, behavior-wise or by any set of behaviors.

Abstract: Traditional artificial intelligence (AI) architectures utilize a centralized approach, in which a single large AI model handles all inference tasks. This approach results in high latency, requires significant computational resources, has difficulty adapting to evolving data, and raises privacy concerns. While hybrid AI architectures have been developed, they suffer from static knowledge bases, limited adaptability, and a lack of continuous learning, which reduces their accuracy. Accordingly, embodiments utilize a hybrid architecture with an iterative local-global model feedback loop for continuous learning during inference. In particular, the local model may escalate inputs to the global model, when it is unable to infer a response with sufficient confidence. The global model may provide a global insight, which the local model may integrate into its response and knowledge base.

Abstract: Conventionally, guardrails for artificial intelligence (AI) agents are static and rigid. As language usage evolves, these guardrails must be manually updated, which has become impractical as the number of AI agents has increased exponentially in recent years. Accordingly, disclosed embodiments provide automated semantic and context-aware expansion of agentic guardrails. In particular, base guardrails may be decomposed into base guardrail elements. The base guardrail elements may be semantically expanded into similar guardrail elements, for which context markers may be generated. New guardrails may be generated by combining these semantically similar guardrail elements with context markers, and these expanded guardrails may be incorporated into the AI agent.

Abstract: A computer-implemented method for calling a REST webservice for interacting with a database that stores database information, the method comprising: presenting to a user a list of pre-existing events, wherein each of the pre-existing events provides a trigger in response to activity in the database; receiving a user selection of one of the presented pre-existing events; receiving a user selection of one or more actions to be taken in response to the selected pre-existing event providing a trigger; and when the selected pre-existing event happens, calling a REST webservice by sending data to a receiver based on the selected pre-existing event and the one or more selected actions.

Abstract: State-of-the-art techniques for detecting personally identifiable information (PII) in data-processing workflows are not scalable, are inefficient, and/or are prone to human error. Accordingly, disclosed embodiments utilize a PII classifier during the design phase or at compile time, for an integration process, to automatically determine the likelihood that each field in the integration code of the integration process is personally identifiable information. Fields identified as likely representing PII data may be automatically excluded from being indexed into the crowd-sourced historical data that are used to train an artificial intelligence (AI) model. This improves operational efficiency, scalability, flexibility, data quality, and accuracy of the AI model, as well as facilitating compliance with data privacy and security regulations, improving trust and adoption, and fostering proactive data management.

Abstract: Currently, technical expertise is required to construct a connector between an integration process and a Java Management extensions (JMX) server. Disclosed embodiments enable automated configuration of a JMX connector, which can be integrated into one or more integration processes. The configuration process may comprise iterating over descriptors of attributes or operations to automatically generate a data schema that can be incorporated into a profile of the JMX connector. The data schema is then used, during execution of the JMX connector, to retrieve attributes of a resource being monitored by the probe layer of JMX or execute an operation available via the application programming interface of the JMX server. The JMX connector enables JMX data to be integrated in the same manner as conventional integration data, and enhances security by enabling JMX data to be confined within the boundaries of the integration platform.