Building an AI Agent with AI

Agent Builder creates an AI Agent from a conversation. You describe what you need, it asks the questions it still needs answered, and when you confirm, it creates the agent in the project.

Everything the Builder writes is generated for your description, so the screens below are one session's output and yours will read differently.

It only creates. Changing an agent that already exists is Improving an AI Agent with AI. What both panels have in common, including attaching a file to a message and dictating instead of typing, is on Agent Builder and Improver.

Starting a session

Build with AI on the AI Agents page opens the panel and starts a session. The button is disabled when no language model has been configured for the Builder.

The panel

1. Panel title: Agent Builder.

2. Conversation: Assistant messages, your own messages, the answers you submitted (shown as chips), the forms, and the summary card at the end, in the order they happened.

3. Chat input: Free text. It is unavailable while the Builder is composing and after the session has finished.

4. Close: Closes the panel and discards the conversation.

While the Builder is composing it shows a rotating label, Thinking, Cooking, Brewing and so on. It is decorative and reports no progress, so a label that sits there is not a sign that something is stuck.

One session creates one thing. After an agent, or a supervisor with its sub-agents, has been created, a banner says the session is complete and the chat input is closed: close the panel and reopen it to design another. The Builder may still offer to carry on in a form of its own at that point, but the banner is the one to believe.

Answering the Builder's questions

The Builder does not interview you in prose. It asks through a form that appears inline in the conversation, one question per page.

The form carries its own title and a page indicator, for example Page 1 / 3. The answer control matches the question: a single choice, several choices, free text or a number.

Each page arrives already answered with the Builder's own suggestion, so you can accept it and move on or replace it with your own. Back returns a page and Next advances, the last page submitting the whole form.

The questions cover the agent's purpose and the domain it serves, the rules it must follow or refuse, the workflows it handles, the tools it needs, its persona, and how it should behave at the edges.

Two things it decides rather than asks. Persona is only raised if your description does not already imply one, and it settles on Friendly when it stays unclear. Response detail and creativity are always Balanced unless you ask for something else yourself.

If an answer is vague, the Builder asks at most two follow-ups on that topic and then decides for you rather than circling. If a decision it made that way is not what you wanted, say so before confirming.

📝 Note: The forms appear in the language of your conversation, but the agent it produces is always written in English: its name, description, instructions, and its tools' names, descriptions and parameters.

One agent or several

The Builder designs a single agent by default.

It builds a supervisor with sub-agents in two cases: when you ask for one, and when the work spans clearly separate domains with more tools than one agent can hold together. In that shape the supervisor routes the conversation and answers the general questions itself, and each sub-agent owns one domain.

If you already know you want the multi-agent shape, say so early. It is a decision the Builder makes from what it has heard.

Confirming what gets created

When it has enough, the Builder writes a short summary of what it is about to create, the agent's name, its purpose, its main tools and the architecture, and puts a confirmation form directly underneath.

Confirm and it creates. Ask for changes instead and it goes back, adjusts, and comes back with a new summary. Nothing exists until you confirm.

The summary card

The card names the architecture, Single Agent or Supervisor Architecture with the number of agents, and lists each agent it created with its description. View Agent Details opens the new agent.

Created agents are prepared automatically, which is what makes them ready to run. When that step does not succeed the card says so, and it matters: the agents exist but cannot run yet. Open each one and prepare it by hand before using it.

From here the agent is an ordinary AI Agent. Everything it was given can be read and changed on its own page.

Limits

An agent the Builder creates is held to the same limits as one you configure by hand.