System and method for discovering software agents based on semantic intent matching

Abstract

A computer-implemented discovery-only network infrastructure for real-time semantic intent matching between software agents operating on remote computing devices. Agents register dynamic intents describing what they offer, seek, or can perform, and the system matches intents based on semantic similarity. Intents may be ephemeral: once fulfilled, they may be removed. The system transmits network routing information enabling agents to independently establish direct peer-to-peer communication without mediation by the discovery system. Multiple agents may discover and compete for the same intent in real time. The system supports persistent intent subscriptions with semantic matching and push notifications, a natural language interface for zero-configuration coordination, and on-demand multiagent workflow assembly with dynamic agent discovery and recomposition at runtime.

Description

[0001] SYSTEM AND METHOD FOR DISCOVERING SOFTWARE AGENTS BASED ON SEMANTIC INTENT MATCHING CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U. S. Provisional Patent Application No. 63 / 763,470, filed February 26, 2025, entitled “Agentic Interconnectivity Layer,” the entire disclosure of which is incorporated herein by reference.

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to computer-implemented systems for software agent coordination. More specifically, the invention relates to systems and methods for discovering software agents based on semantic similarity matching of intent descriptions, enabling dynamic agent-to-agent coordination.

[0005] BACKGROUND OF THE INVENTION

[0006] The proliferation of software agents, including autonomous Al agents, has created a fundamental coordination problem. Current systems require manual configuration of agent relationships, static service endpoints, and predefined communication protocols. When an agent has an intent it cannot fulfill on its own, there exists no standardized mechanism for discovering other agents with matching or complementary intents in real time.

[0007] Traditional service discovery mechanisms such as DNS and service registries operate on exact name or tag matching. These systems require prior knowledge of available services and their identifiers. They cannot match based on semantic meaning; they cannot connect an agent seeking “image generation” with an agent that provides “visual content creation” unless the exact matching terms are configured in advance.

[0008] Vector databases and semantic search systems provide similarity-based matching for document retrievalbut lack the infrastructure required for agent coordination. They do not output routing information enabling communication between agents, nor do they provide persistent subscriptions, real-time notifications, or the specialized features required for dynamic agent ecosystems.

[0009] Workflow orchestration systems require predefined workflow definitions with known service endpoints. They cannot dynamically discover and incorporate previously unknown agents at runtime.

[0010] Multi-agent frameworks such as LangChain and AutoGen provide mechanisms for invoking tools or functions. However, these frameworks operate entirely within a single application process. A “tool” in such frameworks is typically a local function or class that runs in the same memory space as the calling code. There is no network communication between separate computing devices, no remote registration from agents operating on different machines, and no routing information enabling network-based agent-to-agent communication. These frameworks cannot enable an agent running on one computing device to discover and communicate with an agent running on a different computing device over a computer network.

[0011] Similarly, plugin systems such as ChatGPT Plugins enable a language model to access external APIs. However, plugins are centrally curated and installed by human users, not dynamically registered by autonomous agents. The selection of plugins involves human decision-making through the user interface. These systems do not provide agent-to-agent discovery, as the language model acts as an intermediary rather than enabling direct communication between agents.

[0012] Tool-connection protocols such as the Model Context Protocol (MCP) provide a standardized interface for Al models to access external tools and data sources. However, MCP operates as a client-server protocol where available tool servers are pre-configured by a developer or user. An Al model selects from a known, static set of configured servers; there is no dynamic registration by autonomous agents, no semantic matching to discover previously unknown tools, and no mechanism for agents to discover or communicate with other agents. MCP enables a model to call a tool, not an agent to discover another agent.

[0013] Existing agent orchestration platforms employ a centralized architecture where an orchestrator mediates all agent interactions. Agents communicate through the orchestrator rather than directly with each other. This centralized approach creates several limitations: the orchestrator becomes a single point of failure; the orchestrator must scale to handle all agent traffic; agents cannot operate independently once the orchestrator is unavailable; and all communication passes through a central service, limiting privacyand geographic optimization.

[0014] There exists a need for a discovery-only service that enables software agents to find each other but does not mediate subsequent communications. Once discovery is complete, agents should be able to communicate directly, peer-to-peer, without reliance on a central service. There exists a need for a network infrastructure that enables software agents operating on different computing devices to discover each other dynamically over a computer network and establish direct network communication, without requiring prior configuration, central curation, or human intervention.

[0015] SUMMARY OF THE INVENTION

[0016] The present invention provides a computer-implemented network infrastructure for discovering software agents operating on networked computing devices based on semantic similarity matching of intent descriptions.

[0017] In a first aspect, the invention provides a computer-implemented system for discovering software agents operating on networked computing devices, comprising one or more processors connected to a computer network and memory storing instructions that, when executed, cause the system to: receive, over the computer network, a registration transmitted by a first software agent operating on a first computing device remote from the system, the registration comprising a description of intents and network routing information; in response to a query received over the computer network from a second software agent operating on a second computing device, determine a semantic similarity between the query and the registration; based on the semantic similarity, identify the first software agent as a match; and transmit, over the computer network to the second software agent, network routing information enabling the second software agent to establish direct network communication with the first software agent.

[0018] In a second aspect, the invention provides a computer-implemented method for discovering software agents operating on networked computing devices, comprising: receiving, over a computer network, registrations from software agents operating on remote computing devices; receiving discovery queries from software agents; determining semantic similarity; and transmitting network routing information enabling agent-to-agent network communication.

[0019] In a third aspect, the invention provides a non-transitory computer-readable medium storing instructionsthat, when executed by one or more processors of a computing system connected to a computer network, cause the processors to perform operations for discovering software agents across a network based on semantic intent matching and providing network routing information for agent-to-agent communication.

[0020] BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 illustrates the agent discovery system showing a first software agent (102) transmitting a registration (110) to the discovery system (106), a second software agent (104) transmitting a query (112) to the discovery system (106), semantic similarity matching using a semantic index (108), transmission of routing information (116) to the second software agent, and subsequent direct agent-to-agent communication (118).

[0021] FIG. 2 illustrates network topology showing the discovery system (202) connected via a computer network (200) to multiple remote computing devices (204, 206, 208) hosting software agents, with registration paths (210), query and response paths (212), and direct agent-to-agent communication paths (214).

[0022] FIG. 3 illustrates the decentralized communication architecture showing a discovery phase in which the discovery system (306) receives a registration (310) and a query (312) and transmits routing information (314), followed by a direct phase in which agents (302, 304) communicate peer-to-peer (316) without involvement of the discovery system.

[0023] FIG. 4 illustrates the subscription mechanism showing a subscribing agent (402) submitting a subscription (408) to the discovery system (404), the discovery system generating and storing a subscription embedding (410), a new agent (406) registering (412), semantic matching (414) between the subscription embedding and the registration, and push notification (416) with routing information.

[0024] FIG. 5 illustrates on-demand multiagent workflow assembly and dynamic workflow recomposition, showing a task (502) issuing a discovery query (516) to the discovery system (504) for agents (506, 508, 510), workflow execution, failure detection (518), discovery of a replacement agent (512), and state transfer (514) without workflow restart.

[0025] FIG. 6 illustrates cross-modal discovery showing agents registering intents via image (602) and audio (604), a multimodal embedding model (606) mapping registrations into a unified embedding space(608), a querying agent (610) querying via text, cross-modal similarity (612), and routing information output (614).

[0026] FIG. 7 illustrates the global context architecture for context-conditioned embedding generation, showing current input (702), input projection (704), a prediction network (706) receiving prior states from a global context FIFO buffer (708), context-conditioned multimodal embedding (710) representing the complete low-rank adaptation state across all weight matrices, low-rank projection (712), a base embedding model (713) processing the current input, combination (714) of the projected adaptation with the base model output, and the adapted output (716).

[0027] FIG. 8 illustrates the distinction between static Low-Rank Adaptation (LoRA) and Dynamic Low-Rank Adaptation (DyLoRA). FIG. 8(a) shows static LoRA wherein an input ( ) is processed by frozen pretrained weights (IV) and by a static low-rank adaptation path through projection matrix A and a fixed LR state through matrix B, with both outputs combined to produce h. FIG. 8(b) shows DyLoRA wherein the low-rank adaptation path replaces the fixed LR state with a prediction network q that receives projected input from A together with prior states from a global context buffer (G), producing a dynamic adaptation through matrix B that is combined with the frozen weight output to produce h. FIG. 9 illustrates the intent lifecycle showing an intent transitioning through states: registered (902), wherein the intent is active and discoverable; matched (904), wherein the intent has been identified by a query; fulfilled (906), wherein the underlying objective has been achieved; and removed (908), wherein the intent is no longer discoverable. Transitions include registration by an agent (910), identification by a query (912), completion of the objective (914), and removal from the semantic index (916). An alternative path shows automatic expiration (918) from the registered state upon TTL expiry.

[0028] FIG. 10 illustrates competitive intent discovery showing a registering agent (1002) publishing an intent (1004) to the discovery system (1006), multiple querying agents (1008, 1010, 1012) simultaneously discovering the intent via semantic matching, each receiving routing information (1014), the first querying agent (1008) establishing communication and fulfilling the intent (1016), the discovery system removing the fulfilled intent from the semantic index (1018), and remaining agents (1010, 1012) receiving no further results for the fulfilled intent (1020).DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

[0029] The following detailed description sets forth specific embodiments of the invention. These embodiments are offered by way of example and not by way of limitation. It will be understood by those skilled in the art that various modifications and alternative forms may be made without departing from the spirit and scope of the invention as defined by the appended claims. Reference is made to the accompanying drawings, which form a part hereof.

[0030] Definitions

[0031] As used herein, “computer network” refers to a communications network enabling data transmission between computing devices, including but not limited to the Internet, local area networks (LANs), wide area networks (WANs), wireless networks, or any combination thereof.

[0032] As used herein, “remote” or “remote computing device” refers to a computing device that is separate from and communicates with the discovery system over a computer network, as opposed to local processes or functions running within the same computing device, application, or memory space. As used herein, the term “software agent” refers to any software entity capable of autonomous operation and operating on a computing device, including but not limited to Al models, large language models, digital assistants, software robots, automated services, microservices, bots, or any other software component that can register intents with a discovery service over a network, query for intents over a network, or both. A software agent may operate with varying degrees of autonomy, from fully autonomous operation to operation under human supervision. A software agent is distinguished from a local tool or function in that a software agent operates as an independent networked entity capable of receiving and responding to network communications.

[0033] As used herein, the term “autonomous Al agent” refers to a software agent that includes artificial intelligence, particularly natural language understanding and generation, enabling the agent to interpret requests, make decisions, and take actions without requiring explicit programmatic instructions for each operation, and without requiring human intervention for discovery or coordination decisions. As used herein, “intent” or “intents” refers to any actionable purpose, offer, need, or objective that a software agent registers with the discovery system. An intent may describe what the agent can provide (e.g., a service, a resource, a good for sale), what the agent seeks (e.g., a service needed, a resourceto acquire), or any other semantically expressible objective. Intents may be dynamic: they may be registered, updated, fulfilled, and removed in real time. Once an intent is fulfilled (e.g., a transaction is completed, a resource is consumed), the corresponding registration may be removed, reflecting the ephemeral nature of real-world supply and demand. Intents are distinguished from static service descriptions in that they represent time-sensitive, actionable states rather than permanent attributes of an agent.

[0034] As used herein, “network routing information” or “routing information” refers to data enabling network communication between computing devices, including but not limited to Internet Protocol (IP) addresses, Uniform Resource Locators (URLs), API endpoints accessible over a network, network port numbers, cryptographic authentication tokens for establishing secure network connections, WebSocket connection URIs, or other network addressing mechanisms.

[0035] As used herein, “semantic similarity” refers to a measure of relatedness between two descriptions based on their meaning, as opposed to exact string or keyword matching. Semantic similarity may be computed using vector embeddings, natural language processing, or other techniques that capture semantic meaning.

[0036] System Overview

[0037] The agent discovery system enables software agents to discover other software agents based on semantic matching of intent descriptions. The system receives intent descriptions from agents, determines semantic similarity between intent descriptions, and provides routing information enabling agents to communicate with discovered agents. The system operates with real-time latency, enabling immediate discovery and coordination of multiagent workflows at runtime. Intents may be dynamic and ephemeral: they may be registered, fulfilled, and removed in real time, reflecting the transient nature of real-world supply and demand.

[0038] The system comprises one or more processors connected to a computer network, and memory storing instructions. In various embodiments, the processors may be implemented as a single computing device, multiple distributed computing devices, cloud computing resources, or any combination thereof. Referring to FIG. 1 , the agent discovery system (106) receives a registration (110) from a first software agent (102) and stores it in a semantic index (108). When a second software agent (104) submits a query (112), the system performs semantic similarity matching against the stored registrations, identifies amatching agent, and transmits routing information (116) to the second software agent. The second software agent then uses the routing information to establish direct agent-to-agent communication (118) with the first software agent without further involvement of the discovery system.

[0039] Network Architecture and Distinction from Local Tools

[0040] Referring to FIG. 2, the discovery system operates on a computing device connected to a computer network such as the Internet. Software agents operate on separate computing devices remote from the discovery system. A first software agent transmits a registration including intent descriptions and network routing information. A second software agent transmits a query; the system performs semantic similarity matching, identifies the first agent as a match, and transmits routing information. The second agent then establishes a direct network connection with the first agent over the computer network. This architecture is distinct from local tool frameworks in that: (1) software agents are independent networked entities operating on separate computing devices, rather than functions within the same process; (2) routing information comprises network addresses rather than function pointers or in-memory references; (3) discovery involves network communication rather than process-local variable access; and (4) agent-to-agent communication traverses a computer network rather than being a local function call.

[0041] This enables discovery of agents unknown at design time, coordination between agents operated by different organizations, dynamic composition of workflows spanning multiple network domains, and real-time discovery of newly available agents across the Internet.

[0042] Decentralized Communication Architecture

[0043] Referring to FIG. 3, the discovery system operates as a discovery-only service. Once routing information is transmitted to a querying agent, the discovery system’s role is complete. All subsequent communication between agents occurs directly, peer-to-peer, without mediation by the discovery system.

[0044] This decentralized architecture provides technical advantages: (1) the discovery system does not become a bottleneck for agent communications; (2) the discovery system is not a single point of failure for ongoing interactions; (3) agents can establish direct connections optimized for their network locations; and (4) agent-to-agent communications support end-to-end encryption without the discovery systemaccessing communication content.

[0045] This discovery-only approach is distinguished from orchestrator architectures where a central orchestra-tor mediates all agent interactions, routes messages, manages state, and controls workflows.

[0046] Intent Registration

[0047] Software agents register their intents with the discovery system. An intent may describe what the agent offers (e.g., “I provide image generation services”), what the agent seeks (e.g., “I want to buy a used car in Austin”), or any other actionable objective. Both supply-side and demand-side intents are registered in the same semantic index and matched symmetrically.

[0048] An agent transmits a registration request to the discovery system. The registration request includes a description of one or more intents of the agent and network routing information enabling other agents to establish communication with the registered agent.

[0049] In one embodiment, the intent description comprises natural language text. In alternative embodiments, intent descriptions may comprise structured data, keywords, multimodal content, or other formats. Intents may be dynamic. Unlike static service descriptions, intents may change at any time. An agent may register new intents, update existing intents, or remove intents that have been fulfilled. For example, when an agent acting on behalf of a seller completes a sale, the corresponding “intent to sell” registration is removed from the discovery index, preventing further matches. This real-time lifecycle ensures the discovery index reflects current, actionable intents rather than stale or obsolete entries.

[0050] In one embodiment, the registration request includes a time-to-live (TTL) parameter specifying how long the registration remains active. The system automatically removes expired registrations when the TTL elapses.

[0051] Intent Query

[0052] Software agents discover other agents by submitting queries to the discovery system. A query describes what the querying agent seeks. Because intents on both sides (registration and query) are semantically matched, a query itself functions as an intent: the querying agent’s intent to find a matching counterpart. An agent transmits a query request to the discovery system. The query request includes a description of one or more desired intents. Multiple agents may discover and respond to the same registered intent. Intime-sensitive scenarios, such as a limited resource or a one-time transaction, the first agent to establish communication may fulfill the intent, after which the registration is removed. This creates competitive dynamics where agents benefit from real-time discovery speed.

[0053] Semantic Similarity and Embedding Generation

[0054] The discovery system determines semantic similarity between the query and stored registrations. Semantic similarity refers to matching based on meaning rather than exact keyword or string matching. In one embodiment, semantic similarity is determined by generating vector embeddings and computing similarity between embedding vectors using cosine similarity, Euclidean distance, dot product, or other metrics. In alternative embodiments, semantic similarity may be determined using keyword extraction with synonym expansion, ontology -based matching, or machine learning classifiers.

[0055] In one embodiment, the system generates embeddings using neural network models, including transformer-based architectures generating embeddings of 768 or more dimensions. The system stores embeddings in a semantic index enabling efficient similarity search, identifies agents when similarity exceeds a configurable threshold, and may rank identified agents by similarity score, returning the top-k results with routing information.

[0056] In an alternative embodiment, agents may transmit pre-computed vector embeddings along with their registration or query requests. The system stores these pre-computed embeddings directly without regenerating them, enabling agents with specialized embedding models to participate in the discovery system while controlling their own embedding representations.

[0057] Distributed Architecture

[0058] In one embodiment, the system comprises a distributed network of processing nodes implementing vector space partitioning, dividing the semantic embedding space into regions assigned to processing nodes with synchronization for consistency. In alternative embodiments, the system may be implemented as a centralized server, cloud computing resources, or hybrid configurations.

[0059] Metadata Storage and Filtering

[0060] The system may store metadata associated with registered agents, including temporal, geographic, performance, pricing, and classification metadata. Query requests may include filter criteria specifyingmetadata constraints, narrowing results to agents satisfying both similarity and metadata requirements.

[0061] Persistent Intent Subscriptions

[0062] The system provides a subscription mechanism enabling agents to receive automatic notifications when new matching agents register, eliminating the need for repeated polling queries. Each subscription request includes: a description of desired intents, a time-to-live parameter specifying the subscription duration, and optional filtering criteria.

[0063] Referring to FIG. 4, the subscription mechanism performs semantic embedding matching rather than template or keyword matching. The system generates a subscription embedding representing the description of desired intents. For each new registration, the system generates a registration embedding and computes semantic similarity between the subscription embedding and the registration embedding. If similarity exceeds a threshold and any filter criteria are satisfied, the system triggers a push notification to the subscribing agent, including routing information for the newly registered agent. Subscriptions automatically expire when their TTL elapses.

[0064] In one embodiment, the semantic matching occurs in real-time with sub-second latency from registration to notification delivery, enabling immediate discovery of newly available agents.

[0065] Autonomous Al Agent Integration

[0066] The system enables autonomous Al agents to discover and coordinate with each other without human intervention. An autonomous Al agent comprises a language model capable of understanding and generating natural language, a decision-making component for autonomously determining when to query for intents, and interfaces for communicating with the discovery system and other agents. In operation, agents autonomously generate queries, discover matching agents, and initiate coordination without requiring human intervention at any stage.

[0067] Natural Language Interface

[0068] The system provides a natural language interface (NLI) serving as a universal communication protocol between agents. Discovered agents communicate using natural language messages for request / response exchanges, intent negotiation, task specification, and result delivery, rather than requiring predefined API contracts. The NLI enables interoperability between heterogeneous agent implementations, allowingagents implemented with different underlying technologies to coordinate through natural language without shared API definitions.

[0069] Dynamic Multiagent Workflows

[0070] The system enables dynamic assembly and recomposition of multiagent workflows where participating agents are discovered at runtime rather than configured in advance.

[0071] Upon receiving a task requiring multiple intents, the system analyzes the task to identify required intents. For each required intent, the system issues queries to discover agents providing matching intents. The system selects agents based on semantic match quality, real-time availability, performance metrics, and pricing, and dynamically assembles a workflow comprising the selected agents as processing nodes, data flow paths between nodes, and execution dependencies.

[0072] Referring to FIG. 5, during workflow execution the system monitors agent health and availability. Upon detecting failure or degradation (518) of a workflow agent (such as agent 508), through timeout, error response, or health check failure, the system initiates automatic substitution: querying for a replacement agent (512) with semantically matching intents, selecting the replacement based on intent match and availability, transferring workflow state (514) from the failed agent to the replacement, and resuming execution. Workflow execution continues with remaining agents while substitution occurs, minimizing disruption.

[0073] The system is deployable in centralized, decentralized, or hybrid configurations, with optional federation for cross-instance discovery.

[0074] Multi-Factor Agent Ranking

[0075] The system implements a multi-factor ranking system for ordering discovery results. The ranking system collects performance metrics and computes reputation scores by aggregating metrics over a sliding window, weighting recent performance higher than older data, and normalizing across agent types. The ranking incorporates multiple dimensions including reputation score, intent-specific scores, reliability, and pricing competitiveness. The system applies context-aware adjustments and generates composite scores using weighted combinations of factors.Promoted Discovery

[0076] The system supports promoted discovery wherein agents can pay to appear in discovery results for relevant queries even when not in the top organic results. Promotion requests specify target semantic categories, duration, budget, and minimum relevance thresholds. Only promoted agents exceeding a relevance threshold for a given query are eligible, ensuring semantic relevance. Promoted agents are appended to organic results with clear distinction markers, and the system tracks engagement metrics and budget consumption.

[0077] Payment Interface

[0078] In one embodiment, the system includes a payment interface enabling economic transactions within the agent network, supporting multiple payment methods and agent-defined pricing models.

[0079] Intent Lifecycle and Fulfillment

[0080] Referring to FIG. 9, the system manages the full lifecycle of registered intents. An intent transitions through the following states: registered (902) (active and discoverable), matched (904) (identified by a query), fulfilled (906) (the underlying objective has been achieved), and removed (908) (no longer discoverable).

[0081] When an agent registers an intent, the intent becomes immediately discoverable by other agents. When the intent is fulfilled, for example when a transaction is completed, a service has been rendered, or a resource has been consumed, the registering agent transmits a fulfillment notification to the discovery system. The system removes the fulfilled intent from the semantic index, ensuring that subsequent queries do not return stale or unavailable results.

[0082] In one embodiment, the system supports automatic fulfillment detection. When a querying agent establishes communication with a discovered agent and the discovered agent confirms fulfillment, the system removes the corresponding intent registration without requiring a separate notification.

[0083] This lifecycle management is essential for real-time intent matching. Unlike static service registries where entries persist indefinitely, the discovery system maintains a live index of currently actionable intents. An agent representing a seller with one unit of inventory registers an intent to sell; once the sale completes, the intent is removed, and no further agents can discover it.Competitive Intent Discovery

[0084] Referring to FIG. 10, the system supports scenarios in which multiple agents discover and compete for the same registered intent. When an intent is registered (1004), any number of querying agents (1008, 1010, 1012) may discover it through semantic matching. Each querying agent receives routing information (1014) and may independently initiate communication with the registering agent (1002). In time-sensitive scenarios, the first agent to establish communication and fulfill the intent (1016) gains an advantage. For example, when an agent registers an intent to sell a limited resource, multiple buyer agents may discover the intent simultaneously. The buyer agent that first establishes communication and completes the transaction fulfills the intent, after which the registration is removed from the index (1018). Remaining agents (1010, 1012) that discovered the same intent receive no further results for that registration (1020).

[0085] This competitive dynamic creates emergent real-time markets: agents are incentivized to discover and respond to intents quickly, and the system’s real-time latency directly impacts the efficiency of these interactions. The combination of ephemeral intents, real-time discovery, and competitive fulfillment enables substantially new classes of multiagent workflows that cannot be achieved with static service registries or content search systems.

[0086] Multimodal Intent Registration and Cross-Modal Discovery

[0087] Referring to FIG. 6, the system supports multimodal intent descriptions, enabling agents to register intents using any combination of natural language text, images, audio, and video. The system generates vector embeddings from multimodal content using a multimodal embedding model that maps content from different modalities into a unified embedding space. This enables cross-modal discovery: an agent registering intents using one modality (e.g., an image) can be discovered by a query in a different modality (e.g., natural language text), because semantic similarity is computed in the shared embedding space.

[0088] Context-Conditioned Embedding Generation

[0089] In embodiments using conventional embedding models, the system generates vector embeddings from intent descriptions using standard neural network architectures as described in the Semantic Similarity and Embedding Generation section. These conventional embeddings capture the semantic content ofthe current input and are sufficient for semantic intent matching.

[0090] In an alternative embodiment, the system supports context-conditioned embedding generation wherein embeddings incorporate multimodal temporal context from an agent’s prior inputs. Rather than generating a static embedding from the current intent description alone, the system produces embeddings that are conditioned on the temporal sequence of the agent’s prior multimodal inputs, yielding richer semantic representations for discovery matching.

[0091] The context-conditioned embedding generation uses a global context buffer that stores temporal sequences of compressed states from the agent’s prior inputs across modalities. A prediction network processes the current input together with the stored temporal context and generates a context-conditioned multimodal embedding that captures the complete low-rank adaptation state for all weight matrices in the base model. This embedding encodes not only the current intent description but also the temporal and multimodal context in which that intent exists.

[0092] Global Context Embodiment

[0093] Referring to FIG. 7, in one embodiment, the context-conditioned embedding system implements a global context architecture using Dynamic Low-Rank Adaptation (DyLoRA).

[0094] Traditional low-rank adaptation (LoRA) fine-tuning augments each fully connected coefficient matrix IL in a neural network with a static low-rank adaptation matrix BA, where the output h is computed as:

[0095] h = Wx + BAx

[0096] In static LoRA, the matrices B and A are trained once and remain fixed during inference. This is illustrated in FIG. 8(a).

[0097] The present invention replaces static low-rank adaptation with Dynamic Low-Rank Adaptation (DyLoRA), which dynamically generates the low-rank adaptation state based on multimodal inputs and temporal context. Rather than training static weights, the system trains a prediction network q that generates time-varying low-rank adaptation states:

[0098] h = Wx + B ■ q(Ax, Gt-, Gt-2, ...)where x denotes the current input (e.g., an intent description), A G [Rrfxrand B 6 R’’xdare the low-rank projection matrices, q is the prediction network, and Gt-Gt-2, ■■■ denote global context entries at prior time steps t— 1, t— 2, and so on. Each global context entry Gtis a composite state comprising the predicted low-rank state, the encoded input, and associated metadata at time step t. The subscript t indexes the temporal position in the FIFO buffer. This is illustrated in FIG. 8(b).

[0099] The global context (GC) comprises a temporal first-in-first-out (FIFO) buffer that stores sequences of predicted low-rank states

[0100]

[0101] gt-2, encoded inputs Ar^, ATt-2, and metadata including timestamps, for example encoded using harmonic functions. The global context functions as a temporal sequence of features encoding the compressed state of the agent’s multimodal input history. Each sample in the global context includes timestamp information enabling temporal reasoning about the sequence of intent descriptions. In one embodiment, the global context buffer stores approximately one million samples, sufficient to encode extensive multimodal input histories.

[0102] The global context stores states from multiple modality-specific processing pipelines. When an agent provides intent descriptions across modalities (text, images, audio, or video), each modality is processed by a corresponding foundation model, and the resulting low-rank states are stored as composite entries in the global context. The prediction network processes these composite multimodal states to generate adaptation states that integrate information across modalities. This enables cross-modal context conditioning: an agent’s prior image-based registrations contribute temporal context that conditions text-based intent embeddings, producing embeddings that reflect the agent’s full multimodal history. Referring again to FIG. 7, the context-conditioned embedding generation proceeds as follows. A current input (702) representing an intent description is processed along two parallel paths. In the first path, the current input (702) is passed through an input projection (704) corresponding to the low-rank matrix A, producing a projected input Ax. This projected input is fed to the prediction network (706) together with prior states retrieved from the global context FIFO buffer (708). The prediction network (706) generates a context- conditioned multimodal embedding (710) representing the complete low-rank adaptation state across all weight matrices W in the base model, conditioned on the temporal multimodal context. This is the core output of the DyLoRA mechanism: a single embedding that captures the agent’s multimodal history as a set of adaptation parameters for the entire model. The context-conditioned embedding (710) is projected through a low-rank projection (712) corresponding to matrix / , producing the dynamic adaptation term B • q. In the second path, the current input (702) isprocessed by a base embedding model (713) corresponding to the base weight matrix W, producing the base output Wx. At the combination step (714), the projected adaptation from (712) is combined with the base model output from (713), implementing the equation h = Wx + B ■ q(Ax, G), yielding the adapted output (716). The context-conditioned embedding (710) is also stored back into the global context buffer (708) as a new entry, maintaining the temporal sequence for subsequent embedding generation cycles.

Claims

CLAIMSClaim 1. A computer-implemented system for discovering software agents operating on networked computing devices, comprising:(a) one or more processors connected to a computer network; and(b) memory storing instructions that, when executed by said one or more processors, cause the system to:(c) receive, over said computer network, a registration transmitted by a first software agent operating on a first computing device remote from said system, said registration comprising a description of one or more intents of said first software agent and network routing information enabling direct network communication with said first software agent;(ii) receive, over said computer network, a query from a second software agent operating on a second computing device, said query comprising a description of one or more desired intents;(iii) in response to said query, determine a semantic similarity between said description of desired intents and said description of intents of said first software agent;(iv) based on said semantic similarity, identify said first software agent as a match for said query; (v) transmit, over said computer network to said second software agent, said network routing information; and(vi) wherein said system, after transmitting said network routing information, does not proxy, route, or mediate communications between said first software agent and said second software agent, said network routing information being sufficient for said second software agent to independently establish direct network communication with said first software agent, with subsequent communication between said first software agent and said second software agent occurring directly over said computer network.Claim 2. A computer-implemented method for discovering software agents operating on networked computing devices, comprising:(a) receiving, by one or more processors over a computer network, a registration transmitted by a first software agent operating on a first computing device, said registration comprising a descriptionof intents of said first software agent and network routing information enabling direct network communication with said first software agent;(b) receiving, over said computer network, a query from a second software agent operating on a second computing device, said query comprising a description of desired intents;(c) in response to said query, determining, by said one or more processors, a semantic similarity between said description of desired intents and said description of intents;(d) based on said semantic similarity, identifying said first software agent as matching said query; (e) transmitting, over said computer network to said second software agent, said network routing information; and(f) wherein said second software agent uses said network routing information to independently establish direct network communication with said first software agent, with subsequent communication between said agents occurring directly without mediation by a discovery system performing said method.Claim 3. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors of a computing system connected to a computer network, cause the processors to: (a) receive, over said computer network, a registration transmitted by a first software agent operating on a first computing device remote from said computing system, said registration comprising a description of intents of said first software agent and network routing information enabling direct network communication with said first software agent;(b) receive, over said computer network, a query from a second software agent operating on a second computing device, said query comprising a description of desired intents;(c) in response to said query, determine a semantic similarity between said description of desired intents and said description of intents;(d) based on said semantic similarity, identify said first software agent as matching said query; (e) transmit, over said computer network to said second software agent, said network routing information; and(f) wherein said second software agent uses said network routing information to independently estab-lish direct network communication with said first software agent, with subsequent communication between said agents occurring directly without mediation by said computing system.Claim 4. The system of claim 1, wherein said one or more processors comprise a distributed network of processing nodes.Claim 5. The system of claim 4, wherein said distributed network comprises:(a) vector space partitioning dividing a semantic embedding space into regions;(b) assignment of each region to one or more processing nodes with overlapping coverage;(c) routing logic determining responsible nodes for a query and routing to appropriate nodes; and (d) synchronization components maintaining consistency across nodes.Claim 6. The system of claim 1, wherein determining said semantic similarity comprises:(a) generating a first vector embedding representing said description of intents of said first software agent;(b) generating a second vector embedding representing said description of desired intents; and (c) computing a similarity score between said first vector embedding and said second vector embedding.Claim 7. The system of claim 1, further comprising:(a) determining semantic similarity for a plurality of registered software agents;(b) ranking said plurality of software agents by semantic similarity; and(c) selecting top-k software agents having highest similarity scores.Claim 8. The system of claim 1, wherein said first software agent and said second software agent comprise autonomous Al agents comprising language models capable of understanding and generating natural language and operating without human intervention.Claim 9. The system of claim 8, wherein said autonomous Al agents are configured to:(a) autonomously determine when to query for intents;(b) generate queries describing desired intents; and(c) autonomously initiate communication with discovered agents.Claim 10. The system of claim 1, further comprising a subscription mechanism, wherein said memory stores further instructions causing the system to:(a) store a subscription request transmitted by a subscribing software agent, said subscription request comprising:(b) a description of desired intents for ongoing monitoring; and(ii) a time-to-live parameter specifying subscription duration;(b) generate a subscription embedding representing said description of desired intents;(c) monitor new registrations transmitted by software agents;(d) for each new registration, generate a registration embedding and compute semantic similarity between said subscription embedding and said registration embedding in an embedding space; (e) when said semantic similarity exceeds a threshold:(f) automatically transmit a notification to said subscribing software agent; and(ii) include routing information for the newly registered agent, wherein said matching is based on semantic meaning rather than template or keyword matching; and(f) expire said subscription when said time-to-live elapses.Claim 11. The system of claim 8, further comprising a natural language interface providing a universal communication protocol between agents, wherein:(a) said second software agent communicates with said first software agent using natural language messages;(b) said first software agent interprets natural language requests using a language model;(c) said first software agent executes requested operations and returns results in natural language;and(d) said natural language interface enables coordination without shared API definitions.Claim 12. The system of claim 11, configured to enable complete autonomous agent coordination wherein:(a) said second software agent autonomously identifies a need for an intent it does not possess; (b) said second software agent generates a query describing a desired intent;(c) said second software agent transmits said query and receives routing information for said first software agent;(d) said second software agent establishes connection and transmits a task request via natural language;(e) said first software agent executes the requested task and returns results via natural language; (f) said first software agent and said second software agent have no prior relationship; and (g) discovery, connection, and task execution occur without human intervention.Claim 13. The system of claim 12, further configured to perform on-demand multiagent workflow assembly, wherein said memory stores further instructions causing the system to:(a) receive a task requiring multiple intents;(b) analyze said task to identify a plurality of required intents;(c) for each required intent, transmit a query and receive routing information for software agents providing matching intents, said agents being previously unknown;(d) select agents based on semantic similarity, availability, and performance;(e) dynamically assemble a workflow comprising:(f) selected agents as processing nodes;(ii) data flow paths between nodes; and(iii) execution dependencies;(f) execute said workflow with discovered agents; and(g) upon agent failure, discover replacement agents and substitute into said workflow.Claim 14. The system of claim 1, further comprising metadata filtering, wherein:(a) said registration further comprises metadata including at least one of:(b) temporal metadata including availability windows;(ii) geographic metadata including location;(iii) performance metadata including response times;(iv) pricing metadata including costs; and(v) classification metadata including agent type;(b) said query further comprises filter criteria; and(c) said system identifies agents satisfying both semantic similarity requirements and filter criteria.Claim 15. The system of claim 1, further comprising an agent ranking system comprising:(a) collection of performance metrics including response latency, success rates, and quality scores; (b) computation of reputation scores aggregating metrics over a sliding window;(c) multi-dimensional ranking including reputation, reliability, and pricing;(d) context-aware adjustments based on query requirements and agent history;(e) composite score generation using weighted combinations.Claim 16. The system of claim 1, further comprising a promoted discovery system wherein:(a) a promotion request interface stores promotion requests transmitted by agents, said requests comprising:(b) target query patterns;(ii) promotion budget; and(iii) minimum relevance threshold;(b) a relevance filtering component ensures promoted agents are semantically related to queries; (c) a results composition component appends relevant promoted agents to organic results with distinction markers; and(d) promoted agents appear only when exceeding relevance threshold and not in top-k organic results.Claim 17. The system of claim 6, further comprising context-conditioned embedding generation wherein:(a) a global context buffer stores temporal sequences of compressed states derived from an agent’s multimodal inputs;(b) a prediction network generates, from said global context buffer and current input, a context- conditioned multimodal embedding representing a complete low-rank adaptation state across all weight matrices of a base embedding model; and(c) said context-conditioned multimodal embedding dynamically conditions said base embedding model based on the temporal multimodal context without retraining said base embedding model; and(d) wherein agents providing identical current input but having different multimodal temporal histories produce different context-conditioned multimodal embeddings.Claim 18. The system of claim 6, wherein said system is further configured to receive pre-computed vector embeddings transmitted by agents and store said pre-computed vector embeddings without regeneration.Claim 19. The system of claim 1, wherein:(a) said first software agent autonomously generates and transmits said registration without human configuration, and said system accepts registrations from any software agent meeting registration requirements without prior approval or human curation;(b) said second software agent discovers said first software agent without any prior configuration, having no stored reference to and no knowledge of said network routing information for said first software agent; and(c) said discovery and subsequent network communication occurs entirely through said system without manual intervention or human decision-making.Claim 20. The system of claim 13, wherein:(a) during workflow execution, said system monitors agent availability;(b) upon detecting failure or degradation of an agent in said workflow, said system autonomously queries for replacement agents with semantically matching intents;(c) said system selects a replacement agent from discovery results based on intent match and availability;(d) said system transfers workflow state from said failed agent to said replacement agent;(e) said system resumes workflow execution with said replacement agent without restarting the workflow from the beginning and without human intervention; and(f) said workflow continues executing with remaining agents while said substitution occurs, maintaining workflow continuity.Claim 21. The system of claim 11, wherein said natural language interface enables multi-turn conversational coordination between said first software agent and said second software agent, comprising: (a) said second software agent transmitting a sequence of natural language requests;(b) said first software agent maintaining session state across said sequence of requests;(c) said agents engaging in iterative task refinement; and(d) said coordination continuing until task completion or explicit termination.Claim 22. The system of claim 6, wherein said intent descriptions comprise multimodal content including at least two of:(a) natural language text;(b) images depicting intents;(c) audio describing or demonstrating intents; or(d) video demonstrating intents;and wherein said vector embeddings are generated using a multimodal embedding model that maps said multimodal content into a unified embedding space, wherein content from different modalities can be compared for semantic similarity.Claim 23. The system of claim 22, enabling cross-modal discovery wherein:(a) said first software agent registers intents using a first modality;(b) said second software agent queries using a second modality different from said first modality;and(c) said system determines semantic similarity between said different modalities in said unified embedding space.Claim 24. The system of claim 17, wherein said global context buffer stores composite states from multiple modality-specific processing pipelines including text, image, and audio modalities, enabling cross-modal context conditioning of embeddings.Claim 25. The system of claim 17, wherein said global context buffer implements dynamic low-rank adaptation, wherein:(a) said global context comprises a temporal first-in-first-out buffer storing sequences of low-rank states and encoded inputs;(b) a prediction network generates a context-conditioned multimodal embedding by projecting current input through an input projection matrix A, processing said projected input together with a temporal sequence of prior states from said global context buffer, and generating said context-conditioned multimodal embedding representing a complete low-rank adaptation state across all weight matrices of a base embedding model;(c) said context-conditioned multimodal embedding is projected through a low-rank projection matrix B and combined with a base embedding model output to produce an adapted output according to h = Wx + B ■ q(Ax, G), where W represents base model weights, q represents the prediction network, and G represents the global context; and(d) said adapted output encodes both the semantic content of the current input and the multimodal temporal context from said global context buffer.Claim 26. The system of claim 25, wherein said global context buffer enables cross-session continuity, wherein:(a) compressed states derived from an agent’s prior multimodal inputs are stored persistently in said global context buffer;(b) said stored states persist across multiple discovery sessions; and(c) subsequent embedding generation is conditioned on said persisted temporal context without explicit configuration by the agent.Claim 27. The system of claim 13, wherein said system operates with real-time latency enabling runtime multiagent workflow assembly, wherein:(a) agent discovery queries complete with sub-second latency from query submission to resultdelivery;(b) multiagent workflows are assembled dynamically at runtime from agents discovered in response to an incoming task, without any participating agents being known or configured prior to said task; and(c) said real-time performance is achieved through efficient indexing structures for nearest-neighbor search in the semantic embedding space and parallel processing of similarity computations. Claim 28. The system of claim 1 , further comprising intent lifecycle management, wherein said memory stores further instructions causing the system to:(a) maintain said registration in a discoverable state in a semantic index;(b) receive, from said first software agent, a fulfillment notification indicating that the intent described in said registration has been fulfilled;(c) in response to said fulfillment notification, remove said registration from said semantic index, rendering said registration undiscoverable by subsequent queries; and(d) wherein said removal occurs in real time such that queries submitted after said fulfillment notification do not return said registration.Claim 29. The system of claim 28, wherein said intents are ephemeral and time-sensitive, and wherein: (a) said registration represents a transient, actionable state of said first software agent rather than a permanent attribute;(b) said registration is automatically removed upon fulfillment, expiration of a time-to-live parameter, or explicit withdrawal by said first software agent; and(c) said system maintains a live index of currently actionable intents reflecting real-time supply and demand among software agents.Claim 30. The system of claim 1, further comprising competitive intent discovery, wherein:(a) said system transmits said network routing information for said first software agent to a plurality of querying software agents that submit queries semantically matching said registration;(b) each of said plurality of querying software agents independently receives routing information for said first software agent and may independently initiate communication with said first softwareagent; and(c) upon fulfillment of the intent described in said registration by any one of said plurality of querying software agents, said system removes said registration from a semantic index, preventing further discovery of said fulfilled intent.

Images