A
Action
An operation an agent asks the application to perform. A tool call is one common representation of an action.
Agent
A system in which a model can use observations to choose among bounded actions in pursuit of a goal. The surrounding application supplies the loop, tools, policy, state, and stop conditions.
Agentic RAG
A retrieval process in which a model can decide whether to retrieve, rewrite or split a query, inspect results, and retrieve again within explicit limits.
Autonomy
The set of decisions a system may make without approval. Autonomy is a property of the whole system, not the model alone.
C
Checkpoint
A durable record from which an execution can resume without repeating completed effects.
Chunk
A portion of a source document stored as one retrievable unit.
Context
The input available to the model for one generation. It can include instructions, messages, retrieved text, tool definitions, state, and tool results.
Context window
The model-specific limit on tokens considered for a request, including input and generated output according to the provider's accounting rules.
E
Embedding
A numeric representation used to compare items in a learned feature space. Similarity in that space is useful for retrieval, but it is not proof that two texts state the same fact.
Elicitation
An MCP mechanism for obtaining user input needed to continue a request. Modern MCP carries the request through an input_required result and a later retry.
G
Grounding
Connecting a generated answer to supplied evidence or system state. Grounding reduces unsupported generation but does not guarantee correctness.
H
Handoff
A transfer of responsibility and selected context from one agent to another.
Human-in-the-loop
A design that stops at a defined point until a person supplies information or approval.
Human-on-the-loop
A design in which a person supervises execution and can intervene without approving every step.
I
Idempotency
A property that lets repeated requests with the same operation identity produce no additional effect after the first successful application.
Indirect prompt injection
Instructions planted in data that a model later reads, such as a webpage, email, document, or tool result.
M
MCP client
The protocol component that communicates with one MCP server on behalf of a host.
MCP host
The application that coordinates model access, MCP clients, user consent, and security policy.
MCP server
A program that exposes MCP capabilities such as tools, resources, and prompts.
Memory
Application-managed information saved for later use. A model does not gain durable memory merely because a conversation contains earlier messages.
O
Observation
Information returned to the agent after an action or environment query.
Orchestration
The code that controls execution order, branching, budgets, retries, checkpoints, approvals, and termination.
P
Prompt injection
Input crafted to make a model treat untrusted content as instructions that override the application's intended policy.
Prompt
In ordinary LLM use, input instructions and content sent to a model. In MCP, a prompt is a server-exposed, user-controlled template that a client can retrieve.
R
RAG
Retrieval-augmented generation. The application retrieves external material and supplies selected results to a generative model.
Reranker
A model or scoring stage that reorders initial retrieval candidates using a more expensive relevance calculation.
Resource
In MCP, application-controlled context addressed by a URI and read through protocol methods.
S
Structured output
Model output constrained or validated against a declared structure, often a JSON Schema.
Subagent
An agent invoked by another agent or orchestrator for a bounded task. A subagent has its own instructions, context, tools, budget, or some combination of these.
T
Tool
A named operation with declared inputs and outputs that lets an agent inspect or change an external environment.
Tool calling
The process in which a model emits a structured request and application code decides whether and how to execute the corresponding tool.
Trajectory
The ordered sequence of model decisions, tool calls, observations, state changes, and outputs produced during an agent run.
V
Vector database
A data system that indexes vectors and supports nearest-neighbor search, often with metadata filtering. It may complement rather than replace a relational database.
W
Working memory
The selected state and intermediate information available during the current task. It is usually a subset of all stored history.