The builder is one screen with two views of the same agent: Simple and Advanced. Both have the toolbar across the top and a composer for AI changes. Advanced adds a rail of icons on each side and the canvas in the middle, with the composer at its bottom; Simple shows six steps in a column, with the composer at its foot and, on a computer, the canvas beside it. Hover any icon to see its name.

​
Top toolbar

Duplicate and export are also on each row of the Agents list. On narrower screens the toolbar keeps the readiness pill and the next step, and shortens the agent name to fit (hover it to see it in full). Agents, Dashboard and Account show only their icons; hover one to see its name. On the narrowest screens the pill shows the number of issues instead of words.

​
Simple and Advanced

Both views edit the same agent, so a change made in one shows in the other.
  • Simple shows six steps in a column on the left. Who: the agent’s name, personality, language and whether the caller can interrupt, with its model, voice and hearing summed up (Change in Advanced opens them). What it does: each Skill in the order a call meets them, with its instruction, name and When To Use; the first one, which starts the call, also has its greeting. Tools, After the call and Publish are set up in Advanced for now, and their steps take you there. Test opens the Playground. Each step lists what it still needs; a problem the column cannot fix has Change in Advanced, which opens that node’s editor. On a computer the canvas stays next to the column and follows your changes; click a node to bring its step into view. Nothing on the canvas can be moved, connected or deleted in Simple. On a phone the column is the whole screen.
  • Advanced is the full canvas the rest of this page describes: both rails, the palette, the node editor, and moving, connecting or deleting nodes.
An account with no saved agents starts in Simple; everyone else starts in Advanced. Switching with Simple / Advanced is remembered for your account in this browser; a trip to Advanced from a Simple step lasts until you leave the page. Undo, the readiness pill and the next step work the same in both views.

​
Left rail

​
Right rail

Only one right-rail panel is open at a time. Press Escape or the close button to put it away.

​
Canvas

  • Pan by dragging empty space. Zoom with the scroll wheel or a pinch.
  • Select a node by clicking it. Its editor opens on the right. The other nodes stay where they are.
  • Move a node by dragging its header.
  • Delete a selected node or connection with Backspace or Delete, or use the trash icon in a node’s header. Undo brings it back.
  • Bottom-right controls: Auto Arrange tidies the layout, Fit View zooms to show everything.
The composer at the bottom of the canvas talks to the AI builder. Type what you want changed and press Send.

​
Add a node

Open Components. Nodes are grouped: Persona, Speech-to-speech, Conversation, Tools, Post Processing and Advanced. Use the search box to filter by name. Searching for “realtime” also finds the speech-to-speech nodes, which older flows and exports call Realtime Agent and OpenAI Realtime, Gemini Realtime and so on. To add a node, do any of these:
  • Click the row. The node is placed near the related node and selected.
  • Press Enter on a focused row. Same as clicking.
  • Drag the row onto the canvas to choose the position yourself.
To connect two nodes, drag from a dot on one node to a dot on another. The builder only lets compatible dots connect and routes the line to the right input for you. If a connection is refused, the two nodes cannot be joined that way. The Node reference lists what each node accepts.

​
Node editor

Selecting a node opens its settings on the right. On a Persona, each section (Model, Voice, Hearing) shows its provider and model in full on its own line, for example “OpenAI / gpt-5.4-mini”. Fields are in the order you are most likely to need them. The first field of a Skill is Instruction, the one field it cannot run without. Required fields are marked. While any are empty the editor says so at the top (“1 required field is empty”) and the readiness pill reports the same. Less common settings sit under More settings. Click it to expand. Changes apply immediately. There is no save button.

​
Agents list

Open the Agents menu from the toolbar. Each row shows the agent’s name and when it was last changed. From a row you can:
  • Open the agent on the canvas (click or Enter).
  • Rename it (pencil icon, or double-click the name).
  • Duplicate it to start a variant.
  • Export it as an agent file (a .json file you can import again).
  • Delete it. You are asked to confirm.
+ New starts an empty agent. The search box filters by name. Your agent is saved automatically a moment after each change and again when you leave the page. The address bar carries the agent id as ?agent=, so you can bookmark an agent or send the address to a teammate in the same workspace.

​
Version history

Every change creates a version named after what you did, for example “Add Skill” or “Edit Persona name”. AI builder changes are named after the request. Open Version history on the left rail to see the newest 50, newest first. The live version is always listed, even when it is older.
  • Click the restore icon (Restore this version), then Restore to put the canvas back to that state, or Cancel to keep the canvas as it is. Later versions stay in the list, so you can restore forward again. Your next edit after a restore starts a new version.
  • The version that is currently published is labelled Live. Earlier published versions are labelled Previously published.
  • Versions belong to the agent. Switching agents shows that agent’s history.

​
Keyboard shortcuts