AI Agent Glossary | AIRundown

A

A2A Agent Card

An A2A Agent Card is a JSON metadata document that describes an A2A server’s identity, capabilities, skills, service interfaces, and security requirements.

A2A Artifact

An A2A Artifact is an output generated by a remote agent for an A2A Task and composed of one or more content parts.

A2A Message

An A2A Message is a communication turn between an A2A client and remote agent, identified by a role and composed of one or more content parts.

A2A Task

An A2A Task is a server-managed, stateful unit of work with a unique identifier and a defined lifecycle in the Agent2Agent Protocol.

Action

An action is an operation an AI agent chooses to perform to make progress toward a goal, such as calling a tool, querying data, changing a system, or replying to a user.

Agent Graph

An agent graph represents an agentic process as connected nodes and edges, where nodes perform work and edges define possible transitions.

Agent Handoff

An agent handoff transfers active responsibility for a task or conversation from one agent to another, along with the context and state needed to continue.

Agent Instructions

Agent instructions are the rules and guidance that define an AI agent’s role, priorities, constraints, tool-use behavior, output expectations, and conditions for approval or escalation.

Agent Loop

An agent loop is the repeated cycle in which an AI agent evaluates its goal and current state, chooses an action, observes the result, updates state, and decides what to do next.

Agent Memory

Agent memory is the capability of an AI agent to retain and retrieve useful information so it can use prior interactions, experiences, or knowledge in later decisions.

Agent Router

An agent router is a component that classifies a request or task and directs it to the most suitable agent, tool, or workflow.

Agent State

Agent state is the structured record of an AI agent’s current task situation, including its goal, progress, decisions, tool results, pending work, and conditions needed to continue or stop.

Agent Trajectory

An agent trajectory is the ordered sequence of states, decisions, actions, and observations produced while an AI agent works through a task.

Agent Workflow

An agent workflow is a structured sequence or graph of steps through which an AI agent performs, coordinates, and completes work.

Agent2Agent Protocol (A2A) is an open protocol that enables independent AI-agent systems to discover capabilities, communicate, and collaborate on tasks without exposing their internal implementation.

Agentic AI

Agentic AI is AI designed to pursue goals by deciding what to do, taking actions, observing results, and adapting its next steps with limited or substantial autonomy.

AI Agent

An AI agent is a software system that uses an AI model to understand a goal, make decisions, take actions through tools, observe results, and continue working until it reaches an outcome or stopping condition.

API

An API, or application programming interface, is a defined way for software systems to request data or actions from one another.

Autonomy

Autonomy is the degree to which an AI agent can choose and execute actions, manage intermediate steps, and respond to results without requiring human direction at every stage.

C

Chunking

Chunking is the process of dividing large content into smaller units that can be embedded, retrieved, and placed into a model's context.

Context

Context is the information available to a model or AI agent for understanding the current situation and producing its next response or decision.

Context Window

A context window is the amount of tokenized information a model can consider during a single inference.

D

Delegation

Delegation is the act of assigning a bounded task or subgoal to another agent while the delegating component retains responsibility for the larger objective.

Dense Retrieval

Dense retrieval finds relevant items by comparing learned embedding vectors, allowing semantically related content to match even when exact words differ.

E

Embedding

An embedding is a learned numeric representation of data, such as text or an image, designed so that items with related meaning are positioned near one another in a vector space.

Environment

An environment is the external context in which an AI agent operates, including the users, applications, data, tools, rules, and changing conditions it can observe or affect.

Episodic Memory

Episodic memory is stored information about specific past events or interactions, including what happened, when it happened, and what outcome followed.

Evaluation

Evaluation is the systematic measurement of an AI agent's outputs, decisions, actions, and task outcomes against defined criteria.

F

Failure Recovery

Failure recovery is the process of restoring safe, useful agent execution after an error, interruption, or unsuccessful step.

Function Calling

Function calling is a structured interaction in which a model identifies a predefined function and returns arguments for application code to validate and execute.

G

Goal

A goal is the desired outcome an AI agent is expected to achieve. It guides planning, action selection, progress evaluation, and the decision to continue or stop.

Grounding

Grounding is the practice of connecting an AI system's output or decision to relevant evidence, trusted data, rules, or observable real-world state.

Guardrail

A guardrail is a control that constrains, validates, blocks, or escalates an AI agent's inputs, outputs, decisions, or actions.

H

Human-in-the-Loop (HITL)

Human-in-the-loop (HITL) is a system design in which a person reviews, supplies input, approves, corrects, or takes over at selected points in an AI agent’s workflow.

Hybrid Search

Hybrid search combines two or more retrieval methods—commonly sparse keyword search and dense vector search—to rank results using both exact terms and semantic similarity.

I

Idempotency

Idempotency is the property that repeating the same operation has the same intended effect as performing it once.

Inference

Inference is the process of running a trained machine-learning model on an input to produce an output, such as generated text, a prediction, or a structured tool request.

K

Knowledge Base

A knowledge base is an organized collection of information that people or software can search, retrieve, and use to answer questions or support decisions.

L

Large Language Model (LLM)

A large language model (LLM) is a neural network trained on large amounts of text and other data to predict and generate token sequences, enabling language understanding and generation tasks.

Latency

Latency is the elapsed time between starting an AI-agent request or operation and receiving the relevant response or completion.

Long-Term Memory

Long-term memory is information an AI agent retains for future use across separate sessions, conversations, or tasks.

M

MCP Client

An MCP client is the protocol component an AI application creates to connect to and communicate with one MCP server.

MCP Prompt

An MCP prompt is a named prompt template that an MCP server exposes for clients to discover, retrieve, and customize with arguments.

MCP Resource

An MCP resource is URI-identified data that an MCP server exposes for a client to read and use as context.

MCP Server

An MCP server is a program that exposes tools, resources, prompts, or other capabilities to an AI application through the Model Context Protocol.

MCP Tool

An MCP tool is a server-exposed operation that an AI model can discover and request through the Model Context Protocol.

MCP Transport

An MCP transport is the binding that carries MCP messages between a client and server, defining framing and delivery without changing the messages’ protocol meaning.

Metadata Filtering

Metadata filtering restricts retrieval candidates using structured attributes such as date, source, region, document type, owner, or access permission.

Model Context Protocol (MCP)

Model Context Protocol (MCP) is an open protocol that standardizes how AI applications connect to external data sources, prompts, tools, and other capabilities.

Multi-Agent System

A multi-agent system is a system in which two or more AI agents coordinate, collaborate, or divide responsibilities to pursue a goal.

O

Observability

Observability is the ability to understand and diagnose an AI agent's behavior from execution signals such as traces, logs, metrics, events, costs, and outcomes.

Observation

An observation is information an AI agent receives after inspecting or acting on its environment, such as a tool result, error, user reply, sensor reading, or changed system state.

Orchestrator

An orchestrator is the component that coordinates work across agents or workflow steps by routing tasks, tracking progress, managing dependencies, and combining results.

P

Persistent Memory

Persistent memory is agent information saved outside the current runtime so it remains available after a session, process, or device restarts.

Planning

Planning is the process by which an AI agent organizes future tasks, dependencies, actions, tools, and checkpoints to reach a goal.

Prompt

A prompt is the input given to a language model to guide its response, often containing a request, instructions, examples, data, or other context.

Prompt Injection

Prompt injection is an attack or manipulation that places untrusted instructions in an AI system’s input to make the model ignore trusted guidance, reveal data, or take unintended actions.

Q

Query Rewriting

Query rewriting is the process of transforming a user's request into one or more clearer, more specific, or more searchable queries while preserving the original intent.

R

Reasoning

Reasoning is the process of interpreting a goal and available information, evaluating possibilities, and deciding what conclusion or next action best fits the situation.

Reflection

Reflection is the process through which an AI agent evaluates its progress, actions, or output, identifies problems, and decides how to improve or continue.

Replanning

Replanning is the process of revising an existing plan when new observations, failures, or changed constraints make the current path ineffective or invalid.

Reranking

Reranking is a second-stage relevance process that reorders an initial set of retrieved results so the most useful items appear first.

Retrieval

Retrieval is the process of searching an external information source and selecting items that are relevant to a current query, goal, or decision.

Retrieval Pipeline

A retrieval pipeline is the sequence of stages that turns an information need into selected evidence, often including query preparation, search, filtering, ranking, and context assembly.

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) is a technique that retrieves relevant information from an external source and places it into a model's context before the model generates an answer.

Retry

A retry is another attempt to perform an operation after a previous attempt failed or produced a retryable result.

S

Sandbox

A sandbox is an isolated execution environment that limits what agent-run code or tools can access, change, or communicate with.

Semantic Memory

Semantic memory is stored knowledge about facts, concepts, meanings, and relationships that is not tied to one specific past event.

Short-Term Memory

Short-term memory is information an AI agent retains temporarily for a conversation, session, or limited task period before it expires or is discarded.

Sparse Retrieval

Sparse retrieval finds information using mostly zero-valued feature representations in which exact words or lexical terms receive meaningful weights.

State Machine

A state machine is a system model that defines a finite set of states, the events or conditions that trigger changes, and the allowed transitions between states.

Stopping Condition

A stopping condition is a rule that tells an AI agent when to finish, pause, escalate, or abandon its current task.

Sub-Agent

A sub-agent is an AI agent assigned a bounded part of a larger task by another agent or orchestrator, usually with its own context, instructions, and execution loop.

Supervisor Agent

A supervisor agent is an AI agent responsible for directing, reviewing, and coordinating the work of other agents or sub-agents.

System Prompt

A system prompt is a high-priority set of instructions supplied by an application to define a model or agent's role, behavior, priorities, and boundaries.

T

Task Decomposition

Task decomposition is the process of breaking a complex goal into smaller, manageable tasks or subgoals that can be completed and verified.

Timeout

A timeout is a limit after which an agent operation is treated as incomplete or failed because it did not finish within the allowed time.

Token

A token is a unit of text that a language model processes or generates, such as a word, part of a word, punctuation mark, or other text fragment.

Token Cost

Token cost is the monetary or compute expense associated with processing input and generating output tokens during model use.

Tool Calling

Tool calling is the mechanism by which a model or AI agent requests an external capability by selecting a defined tool and supplying the arguments needed to run it.

Tool Definition

A tool definition is a machine-readable description of an available tool, including its name, purpose, and expected inputs, that helps a model decide when and how to request it.

Tool Permission

A tool permission is an enforceable rule that determines whether an AI agent may access a tool or perform a specific operation with it.

Tool Result

A tool result is the output returned after a requested tool is executed, such as data, a confirmation, an error, or a partial-completion status.

Tool Schema

A tool schema is the structured contract that defines the arguments a tool accepts, including field names, data types, required values, and constraints.

Tool Selection

Tool selection is the decision process through which an AI agent chooses which available tool, if any, best supports its current goal and next step.

Trace

A trace is a structured execution record that links the operations within an agent run, including model calls, tool calls, state changes, timing, and errors.

V

Vector

A vector is an ordered list of numbers that represents a point or direction in a mathematical space; in AI systems, vectors often encode features or meaning for comparison.

Vector Database

A vector database is a data system designed to store vector representations and retrieve the most similar items efficiently.

W

Workflow Orchestration

Workflow orchestration is the coordination of steps, dependencies, state, retries, approvals, and external systems required to execute a workflow reliably.

Working Memory

Working memory is the task-relevant information an AI agent actively maintains and uses while solving a current problem.