Creating agents

An agent is a prepared chat for a task that comes up again and again: writing quotes, answering customer enquiries, drafting minutes. You define once how it works, what it knows and which templates it has to hand. From then on, everyone allowed to use it starts a chat with exactly these settings in one click.

In Custodos, recurring workflows are built as agents. An agent always works in the chat: it replies when someone writes to it and never starts on its own.

1. 1

   Open agents

   Click Agents in the sidebar, then New agent. To start from a ready-made draft, choose the Templates tab instead, see Templates.
2. 2

   Describe it in your own words

   On the Create tab, write what the agent should do. The fields fill themselves as you go.
3. 3

   Check the fields

   On the Configure tab you see and change each field individually.
4. 4

   Test

   In the Preview you try the draft out in a real chat.
5. 5

   Save

   Click Create. Until then nothing is saved and the editor shows Unsaved. An existing agent is saved with Save.

The Create tab is a short chat about the agent itself. You describe the task, who it is for, the tone and the rules. Custodos replies and enters matching values in the fields; the fields that changed are listed under Updated: …. Further messages ask for changes.

This conversation saves nothing. It runs on your workspace's default model and is charged in credits like a chat message. The transcript only lasts while the editor is open; the filled-in fields become part of the agent when you save.

Name and Instructions are required, every other field is optional.

The editor explains the split under Instructions or Brain?. The rule of thumb: the instructions describe how the agent works. A brain holds what it needs to know. Files are templates it fills in during any chat.

You can only pick brains you are allowed to read yourself. Once an agent has brains, only models that can search them can be selected; the editor points this out with With brains, only models that can search them.

Which models exist is set by the admins and depends on the processing region the workspace has chosen. The model picker shows what is available. The same applies to web search: if the workspace does not offer it, the switch does not appear and the agent does not search the web, even if a template intends it to.

More under Models and Web search.

The preview is a real chat with the unsaved draft. The same rules apply as later for everyone who uses the agent: data protection settings, brain permissions, files, tools. Messages in the preview are charged in credits like ordinary chat messages.

On wide screens the preview sits to the right of the editor and can be closed with Hide preview. On narrow screens it is a tab of its own. Start over begins a fresh test chat with the same draft. The preview does not appear in your chat history.

Changes only reach other people once you save. From then on, depending on the field:

In the list under Agents, Start chat opens a chat with the agent. The empty chat shows the conversation starters; one click sends them. The agent's files sit ready in the message field like your own attachments. You see what the model receives and can remove a file before sending.

You can switch the model in the chat. If you have started a chat with the agent but not written anything yet, Start chat reopens that chat instead of creating a second one. If your workspace starts new chats in security mode, that applies to agent chats too.

Only the person who created the agent and the workspace admins can edit or delete it. Admins can only change agents they can see: a colleague's private agent stays hidden from them as well.

Delete agent removes the agent for good. Chats that were started with it are kept.