Glossary

A2A (Agent-to-Agent) : An open protocol that lets one AI agent call another. VirtuAI can expose an agent over A2A so that other agent platforms, such as Gemini Enterprise, can send it work. The agent’s agent card, which describes it to callers, is listed in its integration URLs.

Access log : The record of every time someone opened another person’s transcript, revealed a masked value, or exported conversations. Found under Monitor. See Conversations.

Agent : An AI assistant with a personality, a model, and the channels it answers on. It can also have knowledge bases, tools, skills and sub-agents.

Agent Builder : A built-in assistant that helps workspace admins create and improve agents by conversation.

Agent ID : The stable identifier of an agent. It appears in the agent’s integration URLs, so changing it breaks any channel that points at the old one.

Approval : A prompt that asks the person in a conversation to allow or deny an action before the agent takes it, such as running a shell command. An unanswered approval is treated as a denial after 5 minutes.

Budget : A daily or monthly token or cost limit for a workspace, optionally for one provider or model. When it is reached, a budget either warns or blocks new requests.

Channel : Where people talk to an agent: web chat, browser voice, WhatsApp, Google Chat, Slack, Telegram, or another agent over A2A.

Chunk : A piece of a knowledge base document. Documents are split into chunks so the agent can retrieve only the passages relevant to a question.

Connected account : A user’s own account in an external service, such as Google Workspace, GitHub or HubSpot, authorized once so that tools act on that user’s behalf.

Conversation : One exchange between a person (or a calling system) and an agent, on any channel. Every conversation is listed in Conversations.

Dataset : A set of test cases used to evaluate an agent. Each case has an input and the checks its answer must pass.

Draft : The pending, unpublished changes to an agent with versioning turned on. Users keep talking to the published version until the draft is published.

Evaluation : A run of a dataset against an agent, or against its draft, where every answer is scored. Results show pass rates, failures with their reasons, latency and cost, and can be compared with an earlier run.

Feedback : The thumbs up or thumbs down that web chat users give an answer. Admins review it under Feedback.

Integration key : A secret that voice connections send in the X-API-Key header or api_key parameter when calling an agent. With enforcement on, connections without a valid key are refused. Other channels are verified by their platform instead. See Integration keys.

Integration URLs : The section of an agent’s page with its web chat link, widget snippet, webhook URLs for WhatsApp and Google Chat, and A2A agent card. Every URL belongs to that agent only.

Knowledge base : A collection of documents an agent searches to ground its answers in your content. See Knowledge bases.

MCP server : A server that offers tools through the Model Context Protocol. You can add your own, or pick ready-made ones such as Gmail, Google Drive, BigQuery or GitHub, and choose which of each server’s tools are made available.

Personality : The instructions an agent follows: its goal, tone, rules, and when to hand off to a person.

Predefined tool : A built-in tool that is ready to attach to an agent without any setup of your own.

Provider : The company whose model an agent uses: Google Gemini, Anthropic, OpenAI or OpenCode Go. Gemini and Anthropic models can also be used through Vertex AI.

Publish : Making an agent’s draft live on every channel, as a new numbered version. Earlier versions stay in the agent’s history, and you can roll back to any of them.

Retention : How long conversation transcripts are kept before they are removed automatically. One year by default, set per workspace.

Role : A named set of permissions given to workspace members. Every workspace starts with admin, user and viewer. See Users and roles.

Scorer : A check applied to each answer in an evaluation. Automatic scorers look for required or forbidden text, patterns, valid JSON, the expected tool calls, latency or cost. Judge scorers ask a model to grade the answer.

Sensitive data : Values such as credit card numbers, API keys, passwords and bank accounts that VirtuAI detects in messages and shows masked. Revealing them requires a specific permission and is logged. See Sensitive data.

Skill : A reusable, written procedure that an agent loads when a task needs it, such as how to prepare a weekly report. Skills are shared across a workspace.

Sub-agent : An agent that another agent can delegate part of a task to. The delegation and its result appear in the worklog.

Tool : An action an agent can take, such as looking up an order or calling your API. A tool can be custom, predefined, or come from an MCP server. See Tools.

Versioning : An agent setting that makes edits go to a draft instead of live, so they can be reviewed, evaluated and published deliberately.

Workspace : The container for everything in VirtuAI: agents, tools, knowledge bases, keys, users, conversations and analytics. Nothing is shared between workspaces.

Worklog : The record, next to a transcript, of what the agent did to produce its answer: the tools it called, the sub-agents it delegated to, and the files it produced.