Skip to main content
Use this reference to understand the terms you see while building and sharing an agent. Home opens your current space, where you can choose an agent to work on.

Account, space and agent

Your account is how you sign in. A space is the workspace that holds agents, files and the space’s plan. An agent is an assistant configured for a particular job. For example, you might sign in to your account, open your personal space and build a café information agent inside it. Creating an agent does not create another login.

Instructions and prompt

Agent Instructions explain the agent’s job, who it helps, its style and its boundaries. People also call these instructions a prompt or a system prompt. For a café helper, an instruction might tell it to answer from the café FAQ and refer people with booking-change questions to a staff member. Instructions guide behavior; they do not supply every fact about your café. Telling the agent to refer someone to staff does not give it a tool to contact staff.

Knowledge and source

Knowledge is reference material you add for an agent to use, such as a document, pasted text or saved website content. A source is one item of that material. The café FAQ is a source; the instruction to use that FAQ is a different part of setup. A file must be attached to the intended agent and ready before that agent can use it for knowledge answers.

Library, folder and attachment

Library stores and organizes files and other knowledge items. A folder groups items. Attaching a file connects that source to an agent; attaching a folder covers its eligible contents, including later additions. Merely saving a file in Library does not attach it to every agent. Choose a folder attachment only when the covered material belongs in that agent’s knowledge.

Model and fine tuning

The model is the AI that produces an agent’s replies. The model picker chooses among the available options. Fine tuning here means changing response settings, such as answer length or wording variation. It does not mean uploading facts or creating a newly trained model. You can begin with the draft’s model choice and learn these controls when you have a specific reason to change them.

Welcome message and starter prompt

A welcome message greets someone opening a chat. A starter prompt is a suggested question the visitor can select. Neither is the same as the agent’s instructions. For example, the welcome might invite a café question, a starter might ask about the menu, and the instructions tell the agent how to answer using the FAQ.

Preview and saved change

A preview lets you see or try the current setup. The builder chat preview is a working conversation. Design’s appearance preview can reflect edits before you select Save Changes. Check the save behavior for the particular control. Seeing a preview is not a universal save confirmation. A widget is the chat window a visitor opens on a website. Embedding means adding the generated installation code to that website. A direct link opens the agent’s standalone chat without requiring your own website. The same agent can use more than one sharing method, but the methods have their own access settings.

Draft, published, public and live

A draft is an agent you can continue preparing. Publishing and public access are separate controls. The shared launch card calls the agent Live when both are in the required state. A public sharing method also needs its own prerequisites, such as Direct link for a standalone link. Saving instructions does not publish the agent or create a marketplace listing.

Plan, credits and agent subscription

A plan belongs to a space. Credits measure usage, rather than representing a fixed number of messages. An agent subscription is a separate purchase for access to an individual paid agent. Upgrading the space to Pro is different from subscribing to a creator’s agent. Billing opens your active space’s billing settings. Check the displayed checkout information to understand the particular purchase.

Owner and permissions

The owner controls the space. A space currently contains only its owner; inviting other members and assigning roles are not available. Being able to chat with an agent does not give someone permission to edit it or inspect other visitors’ conversations. See Who can see which conversations for Analytics access. Use How to get started to connect these concepts in a first-agent journey, or Where things are in Alysium to locate a screen.