Glossary

This page defines the words that the MentorAgent documentation uses. Each word has one meaning on every page.

Agent

A model with instructions and tools, run by Agent Framework. "The agent" is the one your application runs: the coordinator, with its specialists and teams.

Coordinator

The main agent. It receives every message from the user, calls your actions, specialists and teams, and writes the reply.

Specialist

An agent with its own instructions and tools, defined by a class with [MentorAgent] or by a YAML file. The coordinator hands requests to it.

Handoff

Passing a request from the coordinator to a specialist, through the route_to_specialist tool and a small router agent.

Router

The small agent, with none of your tools, that chooses which specialist gets a handed-off request.

Team

A group of agents, defined by a class with [MentorTeam], that discuss a request in an Agent Framework group chat before the coordinator answers.

Team member

One agent of a team, defined by a [TeamMember] property with a role and instructions.

Remote agent

An agent on another host, listed in RemoteAgents. The coordinator reaches it over A2A in the same way as a specialist.

[MentorAgent]

The attribute that marks a specialist. Without brackets, MentorAgent is the library.

Action

A public method of one of your services that MentorAgent offers to the model as a tool, because it has [Description] or [MentorAction].

Read-only action

An action that you declare as one that only reads, with ReadOnly on [MentorAction] or with IsReadOnlyAction. MentorAgent takes your word for it. An action that needs confirmation is never read-only.

Tool

Anything the model can call during a turn: your actions, MCP tools, the methods of skills, team tools, and built-in tools such as route_to_specialist and navigate_to.

Tool name

The name the model sees: the method name in snake case, without a trailing Async. So CloseCaseAsync becomes close_case.

Required call

A tool call that the first model call of a turn must make. RequireLookupForRecordQuestions and RequireActionForChangeRequests ask for one when a classifier says that the message calls for it.

Core tools

The tools that tool filtering always sends: navigation, memory, delegation to specialists and teams, skills and UI actions.

Skill

Instructions the agent loads only when a request needs them: a class with [MentorSkill] or a SKILL.md file, turned on with EnableSkills.

Page context

The data that an interactive page gives the agent through IMentorPageContext, such as the record the user has selected.

UI action

An action that runs on the current page, registered by the page through IMentorPageContext.

Turn

One user message and everything the agent does to answer it, model calls and tool calls included.

Member turn

One time a team member speaks in the group chat. MaxIterations limits how many there are.

Role

A role claim of the signed-in user (ClaimTypes.Role), checked against the RequiredRoles of an action.

Confirmation

The question the user answers before an action runs. A [MentorAction] asks for it when it declares RequiresConfirmation = true. With the default RequireConfirmationForWrites, it also asks when it declares neither ReadOnly = true nor RequiresConfirmation. A tool matched by RequiresApproval asks too.

Confirmation box

The box the widget shows with the action, its arguments, and the Confirm and Cancel buttons.

Approval

Agent Framework's name for a confirmation. These pages use it only in API names (RequiresApproval, AutoApprovalRules, /mentor/approve) and for the approvals that hosted MCP servers ask for.

Gated

Said of an action or tool that has a rule: a required role, a confirmation, or both. "Role-gated" means it has RequiredRoles.

Withheld

Not offered to the caller: the action is missing from the tool list, and a call by its name runs nothing.

Refused

The call reaches MentorAgent, the method does not run, and the model receives an error.

Grounding rule

The rule MentorAgent adds to the prompts it writes: state only facts that a tool returned.

User

The person who writes to the agent: a signed-in user, or an anonymous user who has not signed in.

Caller

Whatever sends a request to an endpoint: a user's client, an MCP client or another agent.

Stop

The request that ends the current turn: the widget's Stop button, or POST /mentor/cancel.

Channel

The way a client reaches the agent: the widget, the SignalR hub, Server-Sent Events, the MCP server or the A2A server.

Host

The .NET application that calls AddMentorAgent() and runs the agent: an ASP.NET Core app, or a .NET MAUI app.

Backend

An ASP.NET Core host with Novior.MentorAgent.Server that serves clients running elsewhere: a WebAssembly client, a JavaScript front end or a mobile app.

Client

Whatever sends messages to the agent: the widget, your own UI, an MCP client or another agent.

Front end

The user interface: the widget, or your own UI in React, Angular, Vue or a mobile app.

Widget

The ready-made chat UI of MentorAgent, <ChatWidget />, for Blazor and .NET MAUI apps.

Circuit

The connection that Blazor Server keeps with each browser tab. Each circuit has its own coordinator and conversation.

Hub

The SignalR hub of Novior.MentorAgent.Server at /mentor-hub. A conversation lasts as long as the connection.

Server-Sent Events (SSE)

The /mentor/chat endpoint, which streams one answer as a series of events. Each request is a new conversation.

MCP (Model Context Protocol)

An open protocol that lets AI applications list and call tools. Your app is an MCP server at /mcp, and an MCP client of the servers in McpServers.

A2A (Agent-to-Agent protocol)

An open protocol that lets agents send tasks to each other. Your app is an A2A server at /a2a, and calls the remote agents in RemoteAgents.

Agent card

The JSON document at /.well-known/agent-card.json that describes your agent to other agents: name, description, A2A skills and endpoint.

A2A skill

An entry of the agent card that tells callers one thing your agent can do. MentorAgent lists one for each action without rules.

Task

A request that another agent sends to your A2A server. MentorAgent runs it as one full turn, in a new session, and returns the whole reply.

Orchestrator

A program or agent that splits a job among other agents and sends them tasks, for example over A2A. It is not the coordinator.

.NET MAUI Blazor Hybrid

A .NET MAUI app whose UI is Blazor inside a BlazorWebView. MentorAgent builds for Android and Windows inside it. Windows is the platform that has been run; iOS and Mac Catalyst are not checked.

Model key

The API key of your model provider.

Model provider

The service that runs the model, such as OpenAI, Azure OpenAI or Ollama.

History provider

The ChatHistoryProvider that stores the messages of a conversation. Without one, MentorAgent keeps them in memory.

Path A, Path B

The two ways to give MentorAgent a model. Path A is ChatClient, and MentorAgent builds the agent. Path B is your own AIAgent in Agent.

Embedding model

A model that turns a text into a list of numbers that represents its meaning, so that MentorAgent can compare texts.

RAG (retrieval-augmented generation)

For each message, before the coordinator answers, MentorAgent adds to the prompt the most relevant documents from your own knowledge base.

RAG source

The class that searches your documents for RAG. It implements IMentorRagSource.

Tool filtering

With EnableToolFiltering, the model receives only the actions closest in meaning to the message, not all of them. When none is close enough, it receives the three best-ranked actions that need no confirmation and no role.

Model routing

Choosing, for each turn, between ChatClient and a stronger, more expensive StrongChatClient.

Hosted tool

A tool that runs at the model provider, not in your app, such as web search or the code interpreter. Roles and confirmation do not apply to it.

Classifier

The model that makes MentorAgent's own short decisions: the safety checks, the off-topic check, the facts to remember, and the question behind RequireLookupForRecordQuestions and RequireActionForChangeRequests. You set it in ClassifierChatClient.

Safety check

An extra model call, before the answer, that checks each message for prompt injection, jailbreaks and attempts to extract the system prompt or credentials (EnableSafetyCheck). The output safety check (EnableOutputSafetyCheck) checks the reply before the user sees it.

Off-topic check

A check on whether a message is about your application. RefuseOutOfScope refuses a message that plainly has nothing to do with it, and answers when unsure. HostedToolDomainCheck is stricter: it keeps hosted tools away from a message that is not about your application.

Context window

The amount of text that the model reads at once.

Conversation

The messages between a user and the agent, from the first message until a reset.

Session

The object on the server that holds the state of one conversation (Agent Framework's AgentSession).

Snapshot

The saved state of a conversation. SerializeSessionAsync returns it in clear, so keep it on the server or in ProtectedLocalStorage. GET /mentor/session returns a protected form that only the user who saved it can restore.

History

The messages of the current conversation, which the model receives again at each turn.

Reducer

The component that keeps the in-memory history short. The default, MentorChatReducer, removes whole turns, oldest first, and keeps every tool call with its result.

Memory

Facts about a user that MentorAgent keeps across conversations, with UseMemoryContext.

Prerelease

A version whose number has a suffix, such as -rc.N. dotnet add package installs it only when you pass the version or the --prerelease option.

Pin

Give a package reference one exact version, such as [1.17.0-preview.260804.1].