How do I build an agent by describing it to the AI?

What it does

You write what you want the agent to do in your own words; the AI builds the steps, connects them and fills in their settings. For example, if you write “Show the visitor a message, then ask a question, and if the answer is 3 let them upload a file”, it builds Message, Question, Condition and File upload steps.

The AI knows every step type and every integration you have connected or can connect. When there's something it can't decide on its own (what to do with an integration, which board or list to write to), it asks you.

The agent it builds is always a draft: visitors don't see it until you publish.

Building a new agent by describing it

1. Open the “Agents” page from the menu and press “New Agent”.
2. In the window, switch to the “Create with AI” tab.
3. Write what you want the agent to do and press Enter (Shift+Enter for a new line).
4. If the AI asks something, answer it (see “The AI's questions” below).
5. When the agent is ready it is created as a draft and the editor opens. The conversation continues in the “Ask AI” panel on the left, where you can ask for changes.

Changing an existing agent

In the editor you can write in two places:

• If the agent has no steps yet, the “What should this agent do?” box in the middle of the canvas. When you send it, the “Ask AI” panel opens on the left.
• If it has steps, the “Ask AI” button on the toolbar. For example, you could write “Add a form asking for an email after the greeting”.

The AI sees the agent as it is now (including changes you haven't saved) and changes only what you asked for. The change is applied to the canvas immediately: removed steps fade away, and added or changed steps are outlined for a few seconds.

• If you don't like it, press the undo button on the toolbar (Ctrl+Z / Cmd+Z). Everything the AI changed is undone in one go.
• Changes aren't saved until you press “Save”. Visitors only see them when you press “Publish”.
• The canvas is locked while the AI is working; you can edit again once the result arrives.

When you click a step, its settings open on the other side and the panel stays open. Even if you close the panel, the conversation isn't lost; open it again with “Ask AI” to continue where you left off. To start over, press the “New conversation” button in the panel header (next to the close button).

The “Conversation history” button in the header lists every AI conversation on this agent. You can open your own and continue where you left off; your teammates' conversations you can read but not continue. Everyone who can edit the agent sees the history.

The AI's questions

When the AI asks a question, a card appears at the end of the conversation:

• Options: pick one or more; if allowed, type your own answer in “Other”.
• An integration's actions: for example, for “What should it be able to do with Trello?” Trello's actions are listed (Create card, Update card…). The ones the AI suggests come pre-selected; choose the ones you want and press “Continue”. Actions that can't be undone are marked “Can't be undone” and those that send on your behalf “Sends on your behalf”.
• An action's settings: for example, the board and list where cards are created. The choices come from your connected account. If you leave a field as “Let the agent fill it in the conversation”, the agent decides that value itself while talking to the visitor.
• Connecting: if an integration isn't connected, the card shows a button that opens its page in a new tab (for example “Open the Trello page”). After connecting, come back to the panel and press “I've connected it”.

Every card has a “Let the AI decide” button: you skip the question and the AI chooses something sensible. While a card is open the text box below is hidden and comes back once you answer; if you want something else, start over with “New conversation”.

Example: if you write “Make it a smart system connected to Trello”, the AI first asks what it should do with Trello, then which board and list cards go to; finally it builds an AI Assistant step whose Integrations tab has Trello turned on with the actions you chose. See AI Assistant step.

Adding sources

With the “+” button to the left of the box (“Add a source”) you can add information the AI should rely on when building the agent:

• “Add a file or image”: a form, document, screenshot or sketch. Images, PDFs, Word and text files can be added, up to 10 MB each.
• “Choose from knowledge base”: entries from your knowledge base. Only the title of team-only entries is used.
• “Choose from connected apps”: an integration you want the agent to use.
• “Add a website address”: the page's text is read.

Up to 10 sources can be added to one message. One person can upload at most 60 files per hour. Files upload as soon as you pick them; a spinning mark shows on the label while they upload, and the message can't be sent until uploading finishes.

Worth a look

When the agent is built, the panel may show a “Worth a look” list:

• An integration step that can't be undone or sends on your behalf (for example deleting a card or sending a message): the AI doesn't confirm these steps for you. The step doesn't run until you click it and confirm it in the step panel.
• An integration that isn't connected: connect it from the “Integrations” page. See Integration Action and Tool Call steps.

The “Problems in {n} steps” warnings on the canvas work as usual. We recommend trying the steps the AI built with “Preview” before publishing.

Who can use it, and limits

Building agents with AI is available to the Owner and Admin roles, who can create and edit agents. It doesn't use your credits.

Each workspace has a daily message allowance: 30 messages per 24 hours on the Free plan, and 100 during the free trial and on other plans. Your answers to cards and messages that end in an error don't count. When your allowance runs low, the remaining number is shown above the box; when it's used up, you're told when you can continue.

A message can be at most 4,000 characters. A flow built by the AI has at most 80 steps.

In a conversation that gets very long, you're asked to start a “New conversation” to continue. A newly created agent counts towards your plan's agent limit.

Your data

What you write, the sources you add and the agent's content are sent to our language model provider to build the agent; the conversations and the files you attach are kept with your account. Don't attach files or text containing other people's personal data. The details are in the “Building agents with AI” section of the Privacy Policy.