docs/agents/create.mdx
The Agents page opens with a prompt box. Write the job in a sentence and the agent is written for you: a name, instructions, and the tools it needs from the apps you have already connected.
<CardGroup cols={2}> <Card title="Describe it" icon="wand-magic-sparkles" color="#8143E3"> *"Summarise my unread emails every morning."* You get a drafted agent to review. </Card> <Card title="Start from a suggestion" icon="copy" color="#F59E0B"> The chips under the box — triage support tickets, research a company, enrich a lead — are the same route with the sentence written. </Card> </CardGroup>Drafting needs an AI provider connected and turned on for chat. Without one the page says so and offers to write the agent by hand instead.
<Tip> Ask for what you want done, not for the tools. The draft only picks an action that reads your data unless the sentence asks for something to be sent, posted or filed, so *"tell me when something changes"* gets an agent that reports, not one that emails. </Tip>New agent, beside the view toggles on the list, skips the drafting and opens an empty agent for you to write yourself. It goes to your personal project, since an agent you have not deliberately placed belongs to you; where a platform does not create personal projects, the button asks which project to use. Either way you can move it later.
A blank agent has no model yet, so it opens straight into its editor with that still to pick.
Start typing in the prompt box and New agents go to appears beneath it, naming the project the agent will belong to and letting you change it before you create it. That matters because the project is the boundary: an agent reaches only its own project's connections, flows, tables and knowledge.
If you have one project, there is nothing to choose and the control stays out of the way. See managing an agent for moving one later.
The Configure panel beside the conversation holds everything that changes what the agent does.
<CardGroup cols={1}> <Card title="Instructions" icon="pen-line" color="#8143E3" horizontal> Write down the job, how to decide, and when to stop. </Card> <Card title="Agent Tools" icon="wrench" color="#0EA5E9" href="/agents/tools" horizontal> Give it the app actions, flows, and MCP servers it needs to reach. </Card> <Card title="Knowledge Base" icon="book" color="#EAB308" href="/agents/knowledge" horizontal> Attach the documents and tables it should look things up in. </Card> <Card title="Model" icon="microchip" color="#EC4899" horizontal> Pick the model that answers and the key that pays for it. </Card> <Card title="Output fields" icon="brackets-curly" color="#16A34A" href="/agents/in-flows" horizontal> Name the fields you want back, so a flow can use the result. </Card> <Card title="Max steps" icon="gauge" color="#F59E0B" horizontal> Cap how many actions it can take in one run, so a stuck agent stops. </Card> </CardGroup>The editor has two tabs beside the Configure panel.
<CardGroup cols={2}> <Card title="Edit with AI" icon="wand-magic-sparkles" color="#8143E3"> Describe a change — *"only reply to paying customers"* — and it rewrites the instructions and tools for you. </Card> <Card title="Test" icon="flask" color="#0EA5E9"> Talk to the agent as it is now, and watch which tools it reaches for. </Card> </CardGroup>Edits are a draft until you press Publish. The header says which state you are in: Changes not live yet while you are editing, Live once published, and Needs a model to run if no model is picked. Publishing updates every flow already using the agent on its next run, with nothing to republish.
<Tip> Brief it like a capable new colleague. Say the goal and the edges, not every keystroke. The two people forget most: what "done" looks like, and which calls it should hand back to a human. </Tip>