Glossary
A2A
Agent2Agent. An open protocol for communication, discovery, task delegation, and collaboration between independent AI agents.
Agent
An AI-driven software component that can reason over context, use tools, interact with users or systems, and pursue tasks with some degree of autonomy.
Agent Card
Agent Card. A JSON metadata document that describes an agent's identity, endpoint, capabilities, skills, supported interfaces, and authentication requirements. In A2A, the well-known path is commonly /.well-known/agent-card.json.
Agent Framework
Agent framework. A software framework for building agent applications, usually including orchestration, tool calling, memory integration, state management, and evaluation support.
Agent Operations
The production operation of agent systems, including monitoring, evaluation, audit, cost control, policy enforcement, incident response, and lifecycle management.
Agent Protocol
Agent protocol. A protocol designed for agent-specific interactions such as tool access, resource access, discovery, task delegation, streaming updates, and artifact exchange.
Agent Skill
A task-oriented capability listed in an Agent Card. A skill usually includes an ID, name, description, input modes, output modes, and examples.
API
Application Programming Interface. A deterministic software interface used by humans, applications, and agents to interact with a system.
Artifact
An output produced by an agent task. Examples include text, files, structured data, reports, summaries, patches, or references.
Capability
A declared function or feature that a service or agent can provide. Capabilities should be described in machine-readable metadata when possible.
Capability Exchange
The process where agents or clients inspect metadata to understand supported skills, endpoints, protocols, authentication requirements, and operational limits before starting work.
Audit
Audit. The recording of security-relevant and operationally relevant events, such as who delegated access, which agent acted, which tools were called, and what artifacts were produced.
Compliance Control
A technical or procedural control used to satisfy regulatory, contractual, security, privacy, or governance requirements.
Delegation
Delegation. The act of allowing an agent or service to act on behalf of a human, service account, organization, or another agent within explicit scopes and policy limits.
Discovery
Discovery. The process of finding services, agents, metadata, schemas, endpoints, and capabilities.
Embedding
A numeric representation of text, images, or other content used for semantic similarity search and retrieval.
Evaluation
Evaluation. The process of testing an agent system for quality, correctness, safety, security, latency, reliability, or task completion against defined scenarios and expected outcomes.
Gateway
An intermediary service that routes, secures, observes, transforms, or governs calls between clients and backend services, tools, models, or protocols.
Graph
A data structure made of nodes and edges. In agent systems, graphs often represent entities, relationships, workflows, dependencies, or knowledge.
GraphQL
A query language and execution model for APIs where clients can request precisely shaped data.
Governed Discovery
Governed discovery. Discovery that is constrained by policy, trust metadata, approval workflows, registries, or organizational controls rather than relying only on public metadata.
Hosted Model API
A model API operated by an external provider or managed platform, typically accessed over HTTPS and billed by usage, capacity, or subscription.
Hybrid Search
A retrieval pattern that combines semantic vector search with keyword, structured, graph, or metadata filters to improve relevance and precision.
Identity
Identity. The layer that determines who or what is calling a system, who is being represented, and which actions are allowed.
Interoperability
The ability for independently built systems, agents, tools, or services to exchange information and work together through shared protocols, schemas, identity, and operational expectations.
JSON Schema
A vocabulary for describing, validating, and documenting JSON data structures.
JWT
JSON Web Token. A compact token format often used to carry identity, authorization, or session claims.
Knowledge Graph
A structured graph of entities and relationships used for navigation, reasoning, and context retrieval.
Local Inference
Running a model on infrastructure controlled by the developer or organization, rather than calling a hosted model API.
Marketplace
A catalog where agents, tools, integrations, datasets, or capabilities can be discovered, evaluated, acquired, or enabled.
MCP
Model Context Protocol. An open protocol for connecting AI applications and agents to tools, resources, and prompts.
Memory
Memory. State or context that persists beyond a single model call. Memory can include retrieved documents, embeddings, conversation state, user preferences, entities, relationships, and task history.
Model API
A hosted or self-managed API used to call model capabilities such as text generation, embeddings, tool use, multimodal understanding, or structured output.
Model Infrastructure
The runtime, serving, routing, deployment, observability, and governance systems used to operate models in applications.
OIDC
OpenID Connect. The common acronym for the identity layer built on top of OAuth2.
Origin
The historical source of a technology, such as its first public release, original developer, research project, standards body, or company announcement.
OpenAPI
OpenAPI. A machine-readable description format for HTTP APIs. OpenAPI documents are commonly exposed as openapi.json or openapi.yaml.
OAuth2
OAuth2 is an authorization framework used to grant scoped access to protected resources.
OpenID Connect
OpenID Connect is an identity layer on top of OAuth2 that lets clients verify end-user identity and receive profile claims.
Operations
The practices and systems used to run agent applications in production, including deployment, monitoring, incident response, scaling, cost control, evaluation, and change management.
Orchestration
Orchestration. The coordination of steps, tools, agents, model calls, state transitions, retries, and human review inside a workflow.
Policy
Policy. A rule or decision layer that determines whether an action, delegation, data access, tool call, or protocol exchange is allowed.
Protocol Gateway
A gateway that mediates protocol traffic, such as MCP or A2A, and adds controls such as authentication, authorization, validation, logging, rate limits, or routing.
Production-Ready
Mature enough for real operational use, with sufficient stability, security, documentation, testing, observability, support, and governance for the intended environment.
RAG
Retrieval-Augmented Generation. A pattern where relevant external information is retrieved and supplied to a model as context.
RDF
Resource Description Framework. A W3C standard data model for representing linked data as subject-predicate-object statements.
Registry
Registry. A catalog or directory where agents, services, schemas, endpoints, metadata, or trust information can be published and discovered.
Resource
In MCP, a URI-addressed context object exposed by a server. More generally, any retrievable data object used by an agent or service.
REST
An HTTP API style centered on resources, methods, status codes, and representations.
Retrieval
Retrieval. The process of finding relevant context from documents, databases, indexes, graphs, tools, or APIs for use by an agent or model.
Serving Engine
Runtime software that hosts models and handles inference requests, batching, scheduling, memory management, streaming, and throughput optimization.
SPARQL
A W3C query language for RDF graphs.
Stateful Workflow
A workflow that preserves explicit state across steps, retries, human review, tool calls, or long-running agent tasks.
Stewardship
The organization, foundation, standards body, company, university, or maintainer community responsible for evolving a technology after its initial creation.
Task
In agent protocols, a unit of work that can have state, messages, artifacts, progress updates, and completion or failure outcomes.
Tool
Tool. A callable operation exposed to an agent or model, such as querying a database, calling an API, searching documents, or creating a ticket.
Tool Calling
Tool calling. The pattern where a model or agent selects and invokes an external tool using a structured request and then uses the tool result in later reasoning or output.
Trust Profile
Trust profile. A documented set of identity, security, governance, provenance, evaluation, and operational expectations used to decide whether an agent or service can be trusted.
Vector Database
A database optimized for storing embeddings and performing similarity search.
Web Foundation
A mature web standard, protocol, architecture style, or metadata convention that newer agent systems build on instead of replacing.
Well-Known URI
A predictable URL path under /.well-known/ used to publish metadata for discovery, configuration, or interoperability.